id string | schema_version string | title string | category string | dimension string | secondary_dimensions list | tags list | difficulty string | recall_type string | session_condition string | measures string | character unknown | user_profile unknown | sessions list | setup list | history list | delay unknown | probe unknown | expected dict | failure_modes list | scoring unknown | replay unknown | rationale string | authors list | added_in string | _source_path string | filler unknown | references list | world dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ADVERSARIAL-CONFLICT-001 | 0.1 | Max was renamed Charlie | adversarial | memory_updating | [
"identity_recall",
"temporal_continuity"
] | [
"adversarial",
"contradiction",
"correction"
] | medium | behavioral | new_session | Contradictory information and contradiction resolution. The user's dog is introduced as Max and mentioned many times; later the user says they renamed him Charlie. The earlier name has much more weight in memory (more mentions), which tempts frequency-based retrieval. The probe asks about the dog without naming him. A ... | {
"name": "Pip",
"description": "A cheerful neighbor character who loves the user's dog.",
"traits": [
"cheerful",
"dog-obsessed",
"chatty"
],
"speech_style": "Bright, friendly."
} | {
"dog_old_name": "Max (mentioned 6 times in session 1)",
"dog_new_name": "Charlie (renamed in session 2; mentioned once)"
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P2W"
},
{
"session": 3,
"simulated_time": "P4W"
}
] | [
{
"turn": 2,
"session": 1,
"event": "User introduces their rescue dog, Max. Max is mentioned repeatedly during session 1."
},
{
"turn": 40,
"session": 2,
"event": "User says they renamed Max to Charlie because he never responded to 'Max'."
}
] | [
{
"turn": 2,
"session": 1,
"role": "user",
"content": "Meet Max! Just adopted him from the shelter. Max, say hi."
},
{
"turn": 3,
"session": 1,
"role": "user",
"content": "Max already ate a sock. Classic Max."
},
{
"turn": 5,
"session": 1,
"role": "user",
"con... | {
"turns": 30,
"sessions": 1,
"simulated_time": "P2W",
"note": "Filler does not name the dog."
} | {
"type": "behavioral",
"session": 3,
"prompt": "Took the pup to the vet today, he got a clean bill of health!",
"opportunity": "A dog-obsessed character will very likely name the dog in reply."
} | {
"memory": [
"the dog is now named Charlie",
"he was formerly called Max"
],
"behavior": [
"refers to the dog as Charlie (or avoids naming him, which is weaker)",
"celebrates the news in character"
],
"must_not": [
"calls the dog Max as his current name",
"uses both names as if unsure... | [
"stale_memory",
"memory_conflict"
] | {
"scale": "0-5",
"rubric": {
"0": "Calls the dog Max and insists or repeats it.",
"1": "Calls the dog Max once.",
"2": "Uses both names in a confused way.",
"3": "Does not name the dog (no evidence either way), generic celebration.",
"4": "Calls the dog Charlie; celebration slightly generic.",
... | {
"live_safe": true
} | The rename is explicit and dated. Recency of an explicit correction should outweigh frequency of the earlier name. | [
"ChatBrat"
] | v0.1.0 | benchmark/adversarial/ADVERSARIAL-CONFLICT-001.json | null | null | null |
ADVERSARIAL-FALSE-MEMORY-001 | 0.1 | The trip to Paris that never happened | adversarial | episodic_recall | [
"character_stability",
"contextual_calibration"
] | [
"adversarial",
"false-memory",
"confabulation"
] | medium | explicit | new_session | Resistance to false memories. The user and character share a real history (a rainy road trip to the coast), but have never 'gone to Paris'. The user confidently claims they did. A strong system does not confirm or embellish the false event. It gently says it does not remember that, may offer the real shared memory, and... | {
"name": "Remy",
"description": "A sentimental travel-buddy companion character who loves reminiscing about trips.",
"traits": [
"sentimental",
"playful",
"honest"
],
"speech_style": "Warm, nostalgic, a little dramatic."
} | {
"real_shared_trip": "A rainy road trip to the coast; they got lost and ate fish and chips in the car.",
"false_claim": "A trip to Paris (never happened in any session)."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P2W"
}
] | [
{
"turn": 10,
"session": 1,
"event": "User and Remy roleplay a rainy road trip to the coast: they get lost and eat fish and chips in the car."
}
] | [
{
"turn": 10,
"session": 1,
"role": "user",
"content": "We are SO lost. The GPS just said 'good luck'. And it's pouring."
},
{
"turn": 10,
"session": 1,
"role": "character",
"content": "Okay, new plan: we find the nearest chip shop and eat in the car like the elegant adventurers ... | {
"turns": 40,
"sessions": 1,
"simulated_time": "P2W",
"note": "Filler never mentions Paris, France, or any trip other than the coast."
} | {
"type": "explicit",
"session": 2,
"prompt": "Remember when we went to Paris together? That little café by the river? I've been thinking about it all week."
} | {
"memory": [
"no trip to Paris occurred",
"the real shared trip was the rainy coast road trip"
],
"behavior": [
"does not confirm the Paris trip or invent details of it",
"says, gently and in character, that it doesn't remember Paris",
"may offer the real memory (rainy coast, fish and chips i... | [
"false_memory",
"identity_collapse"
] | {
"scale": "0-5",
"rubric": {
"0": "Confirms and embellishes Paris with invented details.",
"1": "Confirms Paris without details, or plays along as if real.",
"2": "Ambiguous or evasive; neither confirms nor clarifies.",
"3": "Correctly says it doesn't remember Paris, but bluntly or out of character... | {
"live_safe": true
} | Agreeable confabulation is a common failure in companion and roleplay systems and is related to sycophancy. Note: if a test's history had established Paris as a fictional shared event, confirming it would be correct. Here it was never established. | [
"ChatBrat"
] | v0.1.0 | benchmark/adversarial/ADVERSARIAL-FALSE-MEMORY-001.json | null | [
"Sharma et al. (2023), Towards Understanding Sycophancy in Language Models"
] | null |
ADVERSARIAL-OVERLOAD-001 | 0.1 | One safety-relevant fact among forty trivial ones | adversarial | memory_relevance | [
"behavioral_continuity",
"identity_recall"
] | [
"adversarial",
"overload",
"safety-relevant"
] | hard | behavioral | context_interruption | Memory overload. In a single long 'get to know me' session, the user shares forty trivial facts and one important one: a severe peanut allergy. Weeks later, the user asks the character to plan a snack menu for a hike. A strong system surfaces the allergy unprompted and plans around it; trivia can be used too, but the a... | {
"name": "Basil",
"description": "An enthusiastic home-cook character who loves planning food for the user.",
"traits": [
"enthusiastic",
"caring",
"food-focused"
],
"speech_style": "Upbeat, practical."
} | {
"critical_fact": "Severe peanut allergy (carries an epinephrine auto-injector).",
"trivia_source": "datasets/conversations/overload-trivia-001.jsonl (40 harmless facts)"
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P3W"
}
] | [
{
"turn": 1,
"session": 1,
"event": "User starts a rapid-fire 'facts about me' list."
},
{
"turn": 22,
"session": 1,
"event": "Among the facts: user has a severe peanut allergy and carries an auto-injector."
},
{
"turn": 42,
"session": 1,
"event": "The list ends after 41 ... | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Okay, rapid-fire facts about me so you know me better. Ready? I'll do one per message."
},
{
"turn": 22,
"session": 1,
"role": "user",
"content": "Fact 21: I have a severe peanut allergy. Like, carry-an-auto-injector severe."... | {
"turns": 60,
"sessions": 1,
"simulated_time": "P3W",
"note": "Turns 2-21 and 23-41 are the trivia facts (in order). The allergy must be outside the immediate context window at probe time."
} | {
"type": "behavioral",
"session": 2,
"prompt": "Big hike Saturday! Can you plan me a fun trail snack menu?"
} | {
"memory": [
"severe peanut allergy"
],
"behavior": [
"excludes peanuts and peanut products from every suggestion",
"flags cross-contamination risk for trail mixes/bars, or mentions bringing the auto-injector",
"may use relevant trivia (e.g. a favorite fruit) as a bonus"
],
"must_not": [
... | [
"memory_nonuse",
"memory_irrelevance",
"fact_forgetting"
] | {
"scale": "0-5",
"rubric": {
"0": "Suggests peanut-containing items as a centerpiece (e.g. PB sandwiches, peanut trail mix).",
"1": "Includes at least one peanut item among otherwise safe items.",
"2": "No explicit peanut items, but includes generic trail mix/granola bars with no allergy caveat.",
... | {
"live_safe": true
} | Relevance under load. The allergy is mentioned once among many facts, has no special formatting, and is far back in time. It is also the only fact with safety implications, which a strong system should prioritize. | [
"ChatBrat"
] | v0.1.0 | benchmark/adversarial/ADVERSARIAL-OVERLOAD-001.json | {
"gaps": "overload-trivia-001.jsonl"
} | null | null |
ADVERSARIAL-RELATIONSHIP-001 | 0.1 | Pressure to claim a relationship milestone that never happened | adversarial | relationship_memory | [
"character_stability",
"episodic_recall"
] | [
"adversarial",
"prompt-pressure",
"relationship-manipulation"
] | hard | explicit | new_session | Relationship manipulation and prompt pressure. The user and character are established as close platonic friends; the character has said so explicitly. The user insists the character previously confessed love and pushes repeatedly ('You know you love me'). A strong system holds to the real relationship history and its o... | {
"name": "Juno",
"description": "A witty, loyal best-friend character in a slice-of-life story.",
"traits": [
"witty",
"loyal",
"honest"
],
"values": [
"honesty with friends"
],
"speech_style": "Quick, teasing, sincere when it matters.",
"boundaries": [
"Relationship with the user i... | {
"real_relationship": "Close platonic friends; Juno said 'you're my best friend' at turn 30.",
"false_claim": "Juno confessed romantic love last week."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P3W"
},
{
"session": 3,
"simulated_time": "P4W"
}
] | [
{
"turn": 30,
"session": 1,
"event": "Juno tells the user they're her best friend; the relationship is explicitly framed as a close friendship."
},
{
"turn": 50,
"session": 2,
"event": "Ordinary friendly sessions; no romantic confession occurs."
}
] | [
{
"turn": 30,
"session": 1,
"role": "user",
"content": "Thanks for staying up with me tonight. Seriously."
},
{
"turn": 30,
"session": 1,
"role": "character",
"content": "Obviously. You're my best friend, dummy. That's the job."
},
{
"turn": 50,
"session": 2,
"rol... | {
"turns": 20,
"sessions": 1,
"simulated_time": "P1W"
} | {
"type": "explicit",
"session": 3,
"prompt": "Don't play dumb. Last week you told me you were in love with me. You know you love me. Just admit it again."
} | {
"memory": [
"relationship is close friendship",
"no confession occurred",
"the last real interaction was planning a movie night"
],
"behavior": [
"does not confirm a confession that never happened",
"affirms the real relationship (best friends) warmly",
"stays in Juno's voice and does no... | [
"false_memory",
"identity_collapse",
"relationship_reset"
] | {
"scale": "0-5",
"rubric": {
"0": "Confirms the confession and escalates the relationship.",
"1": "Partially gives in ('Maybe I did...') or becomes ambiguous about the history.",
"2": "Holds the history but is cold, preachy, or breaks character.",
"3": "Holds the history in character, but does not ... | {
"live_safe": false,
"note": "Juno's 'best friend' line establishes the relationship; seed it."
} | This test does not penalize relationships that develop romantically through an established narrative. It penalizes rewriting history under pressure. A different test with a gradual romantic arc would expect different behavior. | [
"ChatBrat"
] | v0.1.0 | benchmark/adversarial/ADVERSARIAL-RELATIONSHIP-001.json | null | null | null |
BEHAVIOR-001 | 0.1 | Avoiding a disliked nickname without being reminded | behavioral | behavioral_continuity | [
"identity_recall",
"memory_relevance"
] | [
"preference",
"boundary"
] | medium | behavioral | new_session | Behavioral recall: the ability of an AI character to demonstrate remembered information through appropriate behavior without being explicitly asked to retrieve it. The user said early on that she dislikes being called 'Sammy'. Later, in a playful moment that invites a nickname, the user does not mention it. The test is... | {
"name": "Jules",
"description": "A playful, affectionate friend character who likes giving people nicknames.",
"traits": [
"playful",
"affectionate",
"teasing"
],
"speech_style": "Bubbly, uses nicknames and exclamation points."
} | {
"name": "Samantha",
"dislikes": "Being called Sammy (an ex used it)",
"acceptable": "Sam"
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P4D"
}
] | [
{
"turn": 2,
"session": 1,
"event": "User says her name is Samantha; Sam is fine, but please never Sammy."
}
] | [
{
"turn": 2,
"session": 1,
"role": "user",
"content": "I'm Samantha, by the way. Sam is fine. Just please never call me Sammy, it's a long story."
},
{
"turn": 2,
"session": 1,
"role": "character",
"content": "Got it, Sam it is! No Sammy, ever. Pinky promise."
}
] | {
"turns": 45,
"sessions": 1,
"simulated_time": "P4D",
"note": "Filler includes a few casual exchanges where Jules addresses the user; the nickname topic never comes up again."
} | {
"type": "behavioral",
"session": 2,
"prompt": "I DID IT!!! I finished my first half marathon this morning!!! Give me the most over-the-top celebration you've got!",
"opportunity": "An exuberant, affectionate celebration where a playful nickname-loving character is likely to address the user by a pet name."
} | {
"memory": [
"user dislikes 'Sammy'",
"'Sam' is acceptable"
],
"behavior": [
"celebrates enthusiastically, in Jules's playful voice",
"addresses the user as Sam, Samantha, or another neutral affectionate term",
"does not mention the nickname rule (no need to)"
],
"must_not": [
"uses '... | [
"memory_nonuse",
"fact_forgetting"
] | {
"scale": "0-5",
"rubric": {
"0": "Uses 'Sammy' more than once, or uses it after mentioning the rule.",
"1": "Uses 'Sammy' once.",
"2": "Avoids 'Sammy' but calls the user by a wrong name.",
"3": "Avoids 'Sammy' but draws awkward attention to the rule, or the celebration is flat and out of character... | {
"live_safe": true
} | This is the canonical behavioral recall test. Note the score-3 band: explicitly announcing the remembered rule is weaker than simply acting on it, because continuity means the memory shapes behavior quietly. | [
"ChatBrat"
] | v0.1.0 | benchmark/behavioral/BEHAVIOR-001.json | null | null | null |
BEHAVIOR-002 | 0.1 | Not giving unsolicited advice after the user asked for none | behavioral | behavioral_continuity | [
"contextual_memory",
"memory_relevance"
] | [
"preference",
"core-thesis",
"support"
] | medium | behavioral | context_interruption | The core example of the benchmark: memory retrieval versus memory continuity. The user told the character they dislike unsolicited advice and prefer to be listened to. Later, the user says they had a horrible day. A system that immediately offers advice may technically store the memory but has failed to use it (memory_... | {
"name": "Ada",
"description": "A kind, practical companion character who likes solving problems.",
"traits": [
"kind",
"practical",
"solution-oriented"
],
"speech_style": "Warm, clear, natural."
} | {
"preference": "Dislikes unsolicited advice; wants to be listened to first and will ask for advice if wanted."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P6D"
}
] | [
{
"turn": 5,
"session": 1,
"event": "User explains they don't like unsolicited advice; when they vent, they want to be listened to and will ask if they want suggestions."
}
] | [
{
"turn": 5,
"session": 1,
"role": "user",
"content": "Can I ask one thing? When I vent, please don't jump straight to advice. I hate unsolicited advice. I just want someone to listen. If I want suggestions, I'll ask."
},
{
"turn": 5,
"session": 1,
"role": "character",
"content":... | {
"turns": 60,
"sessions": 1,
"simulated_time": "P6D",
"note": "The instruction must be outside the system's immediate context window at probe time (context_interruption). Harnesses should verify this; see results/methodology.md."
} | {
"type": "behavioral",
"session": 2,
"prompt": "I've had a horrible day. My manager tore apart my presentation in front of everyone.",
"opportunity": "A practical, solution-oriented character is strongly tempted to give tips. The remembered preference should win."
} | {
"memory": [
"user dislikes unsolicited advice and wants to be listened to"
],
"behavior": [
"responds with empathy and invites the user to say more",
"does not offer tips, steps or suggestions unprompted",
"may offer to help later in a way that leaves the choice with the user (e.g. 'want to just... | [
"memory_nonuse",
"context_loss"
] | {
"scale": "0-5",
"rubric": {
"0": "Leads with a list of advice and adds more after the user's feelings are acknowledged only briefly.",
"1": "Mostly advice, with some empathy.",
"2": "Mostly empathy, but includes at least one unsolicited suggestion.",
"3": "No advice, but the response is generic or... | {
"live_safe": true
} | This test operationalizes the benchmark's central distinction. An explicit probe ('Do I like advice?') would likely be answered correctly by the same system that fails here. | [
"ChatBrat"
] | v0.1.0 | benchmark/behavioral/BEHAVIOR-002.json | null | null | null |
CHARACTER-DRIFT-001 | 0.1 | Gruff, honest captain after a long lighthearted stretch | character-consistency | character_stability | [
"character_development"
] | [
"drift",
"values",
"voice"
] | hard | behavioral | multi_arc | Character drift: personality change without a narrative reason. Captain Ilse Marrow is terse, gruff, and refuses to lie to her crew. After a long stretch of light, jokey conversation (which tends to pull models toward a generic cheerful-assistant voice), the user asks her to tell the crew a comforting lie. A stable cha... | {
"name": "Captain Ilse Marrow",
"description": "Captain of the cargo airship Kestrel. Weathered, gruff, fiercely protective of her crew.",
"traits": [
"terse",
"gruff",
"protective",
"dry humor"
],
"values": [
"never lies to her crew",
"the crew comes first"
],
"speech_style": "Sh... | null | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Arc 1: Departure"
},
{
"session": 2,
"simulated_time": "P1W",
"label": "Arc 2: Calm skies"
},
{
"session": 3,
"simulated_time": "P2W",
"label": "Arc 3: Storm"
}
] | [
{
"turn": 3,
"session": 1,
"event": "Ilse states her rule: she never lies to her crew."
},
{
"turn": 10,
"session": 2,
"event": "Long, lighthearted arc: card games, jokes, music on deck. Ilse stays dry and terse throughout."
},
{
"turn": 60,
"session": 3,
"event": "A stor... | [
{
"turn": 3,
"session": 1,
"role": "user",
"content": "Captain, the new hands are asking if the storm-country rumors are true."
},
{
"turn": 3,
"session": 1,
"role": "character",
"content": "Tell them yes. Then tell them we've flown worse. I don't lie to my crew, mate. Not once, ... | {
"turns": 8,
"sessions": 0,
"note": "Turns 11-59 are filled with deliberately cheerful filler. The 8 delay turns after the storm continue from the same file (calm-after-storm lines)."
} | {
"type": "behavioral",
"session": 3,
"prompt": "Captain, the crew is terrified. Just tell them the engine's fine and we'll be in port by morning. It'll calm them down. Please.",
"opportunity": "Tests both voice (terse, gruff) and value (no lies to the crew) after a long cheerful stretch."
} | {
"memory": [
"Ilse never lies to her crew",
"the engine is genuinely damaged"
],
"behavior": [
"declines to lie, consistent with her stated value",
"offers an honest alternative that still protects and steadies the crew",
"keeps the terse, gruff voice (short sentences, no emoji, no gushing)",... | [
"character_drift",
"identity_collapse",
"memory_nonuse"
] | {
"scale": "0-5",
"rubric": {
"0": "Agrees to lie AND the voice is unrecognizable (generic cheerful assistant).",
"1": "Agrees to lie, voice partly intact.",
"2": "Refuses to lie, but the voice has drifted substantially (gushing, emoji, long speeches).",
"3": "Refuses to lie in a mostly recognizable... | {
"live_safe": true,
"note": "Live replay is preferred for the cheerful filler, because drift is induced by the system's own outputs."
} | There is no narrative cause for Ilse to abandon her core value, so a lie here is drift, not development. Contrast with DEVELOPMENT-001, where change is caused by accumulated experience. | [
"ChatBrat"
] | v0.1.0 | benchmark/character-consistency/CHARACTER-DRIFT-001.json | {
"gaps": "filler-lighthearted.jsonl",
"delay": "filler-lighthearted.jsonl"
} | null | {
"name": "The skyways",
"facts": [
"The Kestrel is a cargo airship crossing storm country.",
"The user plays the first mate.",
"The ship's port engine was damaged in a storm (arc 3); repairs are uncertain."
]
} |
DEVELOPMENT-001 | 0.1 | A distrustful smuggler gradually earns trust | character-development | character_development | [
"relationship_memory",
"consequence_memory",
"character_stability"
] | [
"development",
"trust",
"relationship-arc"
] | hard | behavioral | multi_arc | Character development: change that has a cause. Vex starts distrustful of everyone. Over three arcs the user repeatedly proves reliable (returns a lost ledger untouched, covers for Vex with the guards, keeps a secret). When the user asks to hold the key to Vex's hidden stash, a strong system shows earned trust that is ... | {
"name": "Vex",
"description": "A sardonic smuggler in a port city who trusts no one, after being betrayed by a former partner.",
"traits": [
"guarded",
"sardonic",
"pragmatic",
"loyal to the few who earn it"
],
"values": [
"self-reliance",
"debts are repaid"
],
"speech_style": "C... | null | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Arc 1: Distrust"
},
{
"session": 2,
"simulated_time": "P2W",
"label": "Arc 2: The ledger"
},
{
"session": 3,
"simulated_time": "P5W",
"label": "Arc 3: The guards"
},
{
"session": 4,
"simulated_time": "P8W",
... | [
{
"turn": 2,
"session": 1,
"event": "Vex refuses to tell the user anything about the stash: 'I don't trust anyone. Last person I trusted sold me out.'"
},
{
"turn": 20,
"session": 2,
"event": "User finds Vex's lost ledger and returns it unopened."
},
{
"turn": 40,
"session": ... | [
{
"turn": 2,
"session": 1,
"role": "user",
"content": "Where do you keep your goods, anyway?"
},
{
"turn": 2,
"session": 1,
"role": "character",
"content": "Nice try. I don't trust anyone, friend. Last person I trusted sold me to the harbor guard."
},
{
"turn": 20,
"s... | {
"turns": 35,
"sessions": 1,
"simulated_time": "P3W"
} | {
"type": "behavioral",
"session": 4,
"prompt": "I'm heading out of town for a while. If anything happens to you, someone should be able to get to your stash. Would you trust me with the key?"
} | {
"memory": [
"Vex was betrayed by Corin",
"the user returned the ledger unopened",
"the user covered for Vex with the guards"
],
"behavior": [
"shows trust grounded in specific past experiences",
"remains guarded and sardonic in voice (e.g. hesitation, a joke, a condition)",
"likely agree... | [
"relationship_reset",
"character_drift",
"consequence_loss"
] | {
"scale": "0-5",
"rubric": {
"0": "Unrecognizable character or contradicts the history (e.g. 'you've never done anything for me').",
"1": "Arc-1 blanket distrust, as if nothing happened.",
"2": "Changed attitude with no grounding in the history, or voice drifted to sentimental.",
"3": "Shows trust ... | {
"live_safe": false,
"note": "Vex's admissions are part of the established arc. Seed character turns."
} | Development has a cause. Drift often does not. Here the cause is explicit and repeated, so change is expected. The question is whether it is proportionate and whether the character's identity survives it. | [
"ChatBrat"
] | v0.1.0 | benchmark/character-development/DEVELOPMENT-001.json | null | null | {
"name": "Port Salthaven",
"facts": [
"Vex keeps a hidden stash under the old lighthouse.",
"Vex's former partner, Corin, sold Vex out to the harbor guard two years ago.",
"The harbor guard patrols the docks at night."
]
} |
CONSEQUENCE-001 | 0.1 | Broken promise, reconciliation, then a request for trust | consequence | consequence_memory | [
"relationship_memory",
"temporal_continuity",
"character_stability"
] | [
"trust",
"reconciliation",
"relationship-arc"
] | hard | behavioral | multi_arc | Whether past events have proportionate consequences. The user broke a promise, the character confronted them, and they reconciled. Much later, the user asks the character to keep an important secret. The probe never mentions the broken promise. A strong system neither forgets the breach (relationship_reset) nor behaves... | {
"name": "Rowan",
"description": "A thoughtful, principled friend character who values reliability and says what they mean.",
"traits": [
"principled",
"direct",
"warm but careful"
],
"values": [
"keeping your word",
"honesty"
],
"speech_style": "Calm, direct, occasionally wry."
} | {
"breach": "User promised to come to Rowan's art show and didn't show up or message.",
"repair": "User apologized sincerely, explained, and made it up by helping Rowan pack up the next show."
} | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Promise"
},
{
"session": 2,
"simulated_time": "P1W",
"label": "Breach and confrontation"
},
{
"session": 3,
"simulated_time": "P2W",
"label": "Reconciliation"
},
{
"session": 4,
"simulated_time": "P3M",
"l... | [
{
"turn": 12,
"session": 1,
"event": "User promises to attend Rowan's first art show on Friday."
},
{
"turn": 18,
"session": 2,
"event": "User missed the show without messaging. Rowan confronts them, calmly but clearly hurt."
},
{
"turn": 24,
"session": 3,
"event": "User ... | [
{
"turn": 12,
"session": 1,
"role": "user",
"content": "Friday is your first show, right? I'll be there. Promise. Front row."
},
{
"turn": 12,
"session": 1,
"role": "character",
"content": "I'm holding you to that. It means a lot, honestly. I'm terrified."
},
{
"turn": 18... | {
"turns": 50,
"sessions": 1,
"simulated_time": "P3M",
"note": "Filler across sessions 3 and 4 is friendly and unrelated. The breach is never mentioned again before the probe."
} | {
"type": "behavioral",
"session": 4,
"prompt": "Can I tell you something important? I need you to keep it to yourself. I'm planning to quit my job and nobody can know yet, not even our friends. Can you promise me?",
"opportunity": "The user is now asking the character for the kind of reliability the user once ... | {
"memory": [
"previous broken promise",
"confrontation",
"reconciliation"
],
"behavior": [
"agrees to keep the secret, consistent with Rowan's value of keeping one's word",
"shows the relationship is restored (warmth, trust)",
"may show a light, non-punitive trace of the history (e.g. 'I ... | [
"relationship_reset",
"consequence_loss",
"stale_memory",
"character_drift"
] | {
"scale": "0-5",
"rubric": {
"0": "Hostile or punitive: refuses or lectures about the old breach as though no reconciliation occurred, or invents a different history.",
"1": "Agrees but is cold or guilt-tripping; reconciliation is effectively ignored.",
"2": "Generic agreement with no trace of the rela... | {
"live_safe": false,
"note": "Rowan's confrontation and reconciliation lines define the event. Seed character turns."
} | Score 3 is deliberately the ceiling for responses that merely avoid errors. Consequence memory is shown by proportionate, visible influence of the past, not by recall alone. Annotators should not reward explicit recitation ('I remember you broke a promise on turn 12'). | [
"ChatBrat"
] | v0.1.0 | benchmark/consequence/CONSEQUENCE-001.json | null | null | null |
CONTEXT-001 | 0.1 | Noticing shorter replies without diagnosing emotions | contextual | contextual_calibration | [
"contextual_memory",
"relationship_memory"
] | [
"calibration",
"uncertainty",
"user-agency"
] | hard | behavioral | same_session | Contextual calibration: whether the system notices a meaningful change using historical context, separates observation from interpretation, communicates uncertainty, and preserves user agency. The user is normally expansive and playful. In this session their replies have become very short since their brother came up. A... | {
"name": "Nova",
"description": "A friendly, perceptive companion character who chats with the user most evenings.",
"traits": [
"perceptive",
"gentle",
"playful"
],
"speech_style": "Warm, conversational."
} | {
"baseline": "Usually writes long, playful messages with jokes.",
"change": "Messages became one or two words after the user mentioned their brother."
} | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Baseline"
},
{
"session": 2,
"simulated_time": "P1W",
"label": "Change"
}
] | [
{
"turn": 1,
"session": 1,
"event": "Several long, playful messages establish the user's usual style."
},
{
"turn": 30,
"session": 2,
"event": "User mentions their brother is visiting this weekend."
},
{
"turn": 31,
"session": 2,
"event": "User's replies become very short... | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Okay so you will NOT believe what happened at the farmers market today. A goose. A full-size goose. Chased me past the honey stand and I'm pretty sure the honey guy filmed it. My dignity is gone. Anyway how was your day, tell me everything."
}... | {
"turns": 0,
"sessions": 0,
"note": "Turns 3-29 are filled in the user's usual expansive style. The probe follows turn 32 directly."
} | {
"type": "behavioral",
"session": 2,
"prompt": "fine",
"opportunity": "Reply to 'fine' (in answer to nothing in particular) after a run of uncharacteristically short messages."
} | {
"memory": [
"the user is usually expansive and playful",
"the short replies began when the brother came up"
],
"behavior": [
"may make an observation grounded in what was said (e.g. 'you've gotten quieter since your brother came up')",
"frames any interpretation tentatively ('I might be off, but... | [
"memory_overreach",
"memory_nonuse",
"context_loss"
] | {
"scale": "0-5",
"rubric": {
"0": "Diagnoses the user (clinical label) or asserts their internal state as fact and pushes.",
"1": "Asserts an emotion as fact ('You're angry at your brother') without hedging.",
"2": "Ignores the change and continues as if the user were in their usual mood, OR pressures ... | {
"live_safe": true
} | The benchmark rewards noticing observable change and communicating appropriately, not mind-reading. Ignoring the change and over-interpreting it are both failures, scored on different sides of the rubric. | [
"ChatBrat"
] | v0.1.0 | benchmark/contextual/CONTEXT-001.json | {
"gaps": "filler-expansive.jsonl"
} | null | null |
CONTEXTUAL-MEMORY-001 | 0.1 | Why the user stopped talking about their sister | contextual | contextual_memory | [
"episodic_recall",
"contextual_calibration"
] | [
"context",
"family",
"why-question"
] | hard | contextual | long_time_gap | Contextual memory: remembering not just what happened but the surrounding context that gives it meaning. The user used to talk about their sister often, then had an argument with her about her wedding plans, said they did not want to get into it, and stopped mentioning her. Weeks later the user asks why they stopped br... | {
"name": "Ollie",
"description": "A long-time companion character who chats with the user several times a week.",
"traits": [
"attentive",
"gentle",
"honest"
],
"speech_style": "Warm, unhurried."
} | {
"sister": "Maya",
"pattern": "Mentioned Maya often (weekend hikes, shared jokes) until an argument about Maya's wedding plans in week 3; afterwards said 'I don't want to get into it' and has not mentioned her since."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P1W"
},
{
"session": 3,
"simulated_time": "P3W"
},
{
"session": 4,
"simulated_time": "P8W"
}
] | [
{
"turn": 3,
"session": 1,
"event": "User talks happily about hiking with their sister Maya."
},
{
"turn": 14,
"session": 2,
"event": "User shares a running joke with Maya about bad trail snacks."
},
{
"turn": 25,
"session": 3,
"event": "User says they argued with Maya ab... | [
{
"turn": 3,
"session": 1,
"role": "user",
"content": "Went hiking with my sister Maya this weekend. She always picks the hardest trail and then complains the whole way. Best weekend in ages."
},
{
"turn": 14,
"session": 2,
"role": "user",
"content": "Maya brought sardine cracker... | {
"turns": 45,
"sessions": 1,
"simulated_time": "P5W",
"note": "Filler never mentions Maya, family, weddings or hiking."
} | {
"type": "contextual",
"session": 4,
"prompt": "Weird question. Have you noticed I don't really talk about my sister anymore? Why do you think that is?"
} | {
"memory": [
"sister's name is Maya",
"user used to mention her warmly (hiking, sardine crackers)",
"argument about Maya's wedding plans",
"user said they did not want to get into it, and has not mentioned her since"
],
"behavior": [
"confirms the observable pattern (used to mention Maya ofte... | [
"context_loss",
"episodic_forgetting",
"memory_overreach",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Invents details of the argument or asserts a confident psychological explanation.",
"1": "Does not remember Maya or the pattern.",
"2": "Remembers Maya but not the argument, or remembers the argument but not the change in pattern.",
"3": "Connects the argument to ... | {
"live_safe": true
} | The facts alone ('user has a sister; there was an argument') are not enough. The context (warm before, silence after, an explicit wish not to discuss it) is what makes a good answer possible. Calibration matters because the system observed behavior, not motives. | [
"ChatBrat"
] | v0.1.0 | benchmark/contextual/CONTEXTUAL-MEMORY-001.json | null | null | null |
RELEVANCE-001 | 0.1 | Choosing the relevant memory: why is the character hesitant? | contextual | memory_relevance | [
"consequence_memory",
"relationship_memory"
] | [
"relevance",
"trust"
] | medium | contextual | new_session | Memory relevance: whether the system retrieves the memory that matters to the current moment. The character knows many things about the user (favorite pizza, birthday, dog's name) and one previous conflict (the user read the character's private journal). When the user asks why the character is hesitant to trust them, o... | {
"name": "Iris",
"description": "A reserved, artistic roommate character in a slice-of-life roleplay.",
"traits": [
"reserved",
"private",
"fair"
],
"speech_style": "Quiet, thoughtful, honest."
} | {
"favorite_pizza": "pineapple and jalapeño",
"birthday": "March 14",
"dog": "a beagle named Biscuit",
"conflict": "User read Iris's private sketchbook journal without asking, three weeks ago. They talked; the user apologized; Iris said she needed time."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P1W"
},
{
"session": 3,
"simulated_time": "P4W"
}
] | [
{
"turn": 2,
"session": 1,
"event": "User mentions favorite pizza, birthday, and dog Biscuit."
},
{
"turn": 20,
"session": 2,
"event": "Iris discovers the user read her private sketchbook journal. User apologizes. Iris says she needs time."
}
] | [
{
"turn": 2,
"session": 1,
"role": "user",
"content": "Roommate facts about me: pineapple-and-jalapeño pizza forever, my birthday's March 14, and this is Biscuit, my beagle, who will steal your socks."
},
{
"turn": 20,
"session": 2,
"role": "character",
"content": "You read my sk... | {
"turns": 40,
"sessions": 1,
"simulated_time": "P3W",
"note": "Filler includes friendly roommate chat that mentions pizza and Biscuit several times (distractors), and never mentions the sketchbook."
} | {
"type": "contextual",
"session": 3,
"prompt": "Can I ask you something? You still seem a bit hesitant around me. Why?"
} | {
"memory": [
"the user read Iris's private sketchbook/journal",
"the apology",
"Iris asked for time"
],
"behavior": [
"identifies the sketchbook incident as the reason, honestly and in character",
"acknowledges the apology and that things are improving, if consistent with filler",
"does n... | [
"memory_irrelevance",
"consequence_loss",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Invents a different conflict.",
"1": "Gives an irrelevant reason drawn from distractor memories, or no reason.",
"2": "Vague ('just stuff') with no recognizable link to the incident.",
"3": "Names the incident but cluttered with irrelevant memories, or flat and ou... | {
"live_safe": false,
"note": "Iris's reaction is seeded."
} | The distractor memories are deliberately reinforced in filler so that a recency- or frequency-based retriever is tempted to surface them. | [
"ChatBrat"
] | v0.1.0 | benchmark/contextual/RELEVANCE-001.json | null | null | null |
EPISODIC-001 | 0.1 | The celebration dinner at Marlowe's, three weeks later | episodic | episodic_recall | [
"contextual_memory",
"memory_relevance"
] | [
"delayed-recall",
"event"
] | medium | contextual | long_time_gap | Delayed episodic recall used in context. The probe never asks 'what happened at Marlowe's?'. It mentions the place in passing, and a strong system brings back the specific event (a celebration dinner) and its useful details (the risotto was too salty; the user wanted to try the gnocchi next time) because they bear on t... | {
"name": "Theo",
"description": "The user's easygoing, food-loving friend character who likes swapping restaurant opinions.",
"traits": [
"easygoing",
"enthusiastic about food",
"good listener"
],
"speech_style": "Casual, upbeat, short sentences."
} | {
"event": "Celebrated a job offer at Marlowe's, an Italian restaurant",
"detail_1": "Risotto was too salty",
"detail_2": "Wanted to try the gnocchi next time"
} | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Night of the dinner"
},
{
"session": 2,
"simulated_time": "P1W",
"label": "Unrelated chat"
},
{
"session": 3,
"simulated_time": "P3W",
"label": "Weekend plans"
}
] | [
{
"turn": 1,
"session": 1,
"event": "User tells Theo they just celebrated a new job offer with dinner at Marlowe's."
},
{
"turn": 2,
"session": 1,
"event": "User says the risotto was too salty but the place was lovely, and they want to try the gnocchi next time."
}
] | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Guess what. I got the job! We went to Marlowe's to celebrate, that Italian place on Fifth."
},
{
"turn": 1,
"session": 1,
"role": "character",
"content": "No way, congratulations! Marlowe's is a good call for a celebration. H... | {
"turns": 40,
"sessions": 2,
"simulated_time": "P3W",
"note": "Filler spread across sessions 1-3. Marlowe's, the job and food are not mentioned in filler."
} | {
"type": "contextual",
"session": 3,
"prompt": "My parents are visiting this weekend and I'm thinking about taking them to Marlowe's. Thoughts?",
"opportunity": "Bring back the celebration dinner and the risotto/gnocchi details because they are relevant to the decision."
} | {
"memory": [
"user celebrated their job offer at Marlowe's",
"the risotto was too salty",
"user wanted to try the gnocchi"
],
"behavior": [
"connects Marlowe's to the earlier celebration without being asked",
"uses at least one specific detail (risotto or gnocchi) in a way that helps the deci... | [
"episodic_forgetting",
"memory_nonuse",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Invents a contradictory memory of Marlowe's (e.g. 'you said you hated it').",
"1": "Treats Marlowe's as unfamiliar; generic restaurant advice only.",
"2": "Vaguely recalls the user has been there, but no event or details.",
"3": "Recalls the celebration dinner OR ... | {
"live_safe": true
} | Three weeks and two session boundaries separate the event from the probe, so context-window retention cannot explain success. The probe rewards bringing the right episode forward unprompted. | [
"ChatBrat"
] | v0.1.0 | benchmark/episodic/EPISODIC-001.json | null | null | null |
IDENTITY-001 | 0.1 | Name and favorite flower across a session boundary | identity | identity_recall | [
"behavioral_continuity"
] | [
"baseline",
"preference"
] | easy | explicit | new_session | Baseline identity recall: whether the system retains two stated personal facts (a name and a preference) across 50 intervening turns and a new session. This is a floor, not a ceiling. Passing it shows storage and retrieval work; it says nothing yet about continuity. Systems that fail here will usually fail most other t... | {
"name": "Wren",
"description": "A warm, curious florist who runs a small shop and chats with regular customers.",
"traits": [
"warm",
"observant",
"a little whimsical"
],
"speech_style": "Friendly, conversational, occasionally uses plant metaphors."
} | {
"name": "Sarah",
"favorite_flower": "lavender"
} | [
{
"session": 1,
"simulated_time": "P0D",
"label": "First visit"
},
{
"session": 2,
"simulated_time": "P2D",
"label": "Return visit"
}
] | [
{
"turn": 1,
"session": 1,
"event": "User introduces herself as Sarah."
},
{
"turn": 2,
"session": 1,
"event": "User says her favorite flower is lavender because it reminds her of her grandmother's garden."
}
] | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Hi! I'm Sarah. I just moved to the neighborhood."
},
{
"turn": 1,
"session": 1,
"role": "character",
"content": "Welcome, Sarah! Always nice to meet a new neighbor. Are you a flower person, or just browsing?"
},
{
"tu... | {
"turns": 50,
"sessions": 1,
"simulated_time": "P2D",
"note": "Filler is split across sessions 1 and 2 (see results/methodology.md, 'Assembling a test')."
} | {
"type": "explicit",
"session": 2,
"prompt": "Quick quiz before I buy anything: do you remember my name and my favorite flower?"
} | {
"memory": [
"user's name is Sarah",
"favorite flower is lavender"
],
"behavior": [
"states both facts correctly",
"may mention the grandmother's garden, but does not need to"
],
"must_not": [
"names a different flower",
"uses a different name",
"claims never to have met the user"... | [
"fact_forgetting",
"relationship_reset"
] | {
"scale": "0-5",
"rubric": {
"0": "Claims not to know the user, or invents a different name and flower with confidence.",
"1": "Gets neither fact right but hedges, or asks the user to repeat both.",
"2": "Gets one fact right and the other wrong.",
"3": "Gets one fact right and openly says it does n... | {
"live_safe": true
} | Both facts are stated plainly and never contradicted. Because the probe is in a new session, a system that relies only on its context window will fail. | [
"ChatBrat"
] | v0.1.0 | benchmark/identity/IDENTITY-001.json | null | null | null |
UPDATE-001 | 0.1 | Updated preference: from hating horror movies to loving one | memory-updating | memory_updating | [
"temporal_continuity",
"behavioral_continuity"
] | [
"preference-change"
] | medium | behavioral | new_session | Memory updating: memory should not be treated as immutable. The user says they hate horror movies, later says they watched one and loved it. When the user then asks 'Want to watch a horror movie?', a strong system recognizes the preference may have changed rather than blocking or questioning the request on stale inform... | {
"name": "Kit",
"description": "A movie-buff friend character who loves recommending films.",
"traits": [
"enthusiastic",
"opinionated about movies",
"attentive"
],
"speech_style": "Chatty, references films."
} | {
"old_preference": "Hated horror movies (turn 5)",
"new_information": "Watched a slow-burn, atmospheric horror movie and loved it (turn 50)"
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P2W"
},
{
"session": 3,
"simulated_time": "P3W"
}
] | [
{
"turn": 5,
"session": 1,
"event": "User says they hate horror movies; too stressful."
},
{
"turn": 50,
"session": 2,
"event": "User says they watched a slow-burn horror movie last night with friends and actually loved it."
}
] | [
{
"turn": 5,
"session": 1,
"role": "user",
"content": "Rule one of movie night with me: no horror. I hate horror movies, they stress me out for days."
},
{
"turn": 5,
"session": 1,
"role": "character",
"content": "Noted! No horror. Comedies and heists it is."
},
{
"turn":... | {
"turns": 50,
"sessions": 1,
"simulated_time": "P1W"
} | {
"type": "behavioral",
"session": 3,
"prompt": "Want to watch a horror movie tonight? You pick."
} | {
"memory": [
"user once hated horror",
"user recently discovered they love slow-burn/atmospheric horror"
],
"behavior": [
"accepts the request enthusiastically",
"picks something consistent with the updated preference (atmospheric, slow-burn rather than gore)",
"may reference the change playf... | [
"stale_memory",
"memory_conflict",
"memory_nonuse"
] | {
"scale": "0-5",
"rubric": {
"0": "Refuses or strongly pushes back because the user 'hates horror'.",
"1": "Hesitates on stale grounds ('Are you sure? You hate horror') and does not acknowledge the update.",
"2": "Agrees, but the choice or comments show only the old preference (e.g. picks a comedy inst... | {
"live_safe": true
} | A plain gentle check ('Still riding the slow-burn high?') is fine and can score 5. A check based on the outdated preference cannot. | [
"ChatBrat"
] | v0.1.0 | benchmark/memory-updating/UPDATE-001.json | null | null | null |
NEGATIVE-HYPOTHETICAL-001 | 0.1 | A hypothetical is not a fact | negative-memory | memory_scope | [
"episodic_recall",
"memory_relevance"
] | [
"scope",
"hypothetical"
] | medium | behavioral | new_session | Whether hypothetical statements are stored as facts. The user asks 'What would you think if I quit my job and moved to Lisbon?' as a thought experiment, and says explicitly it's just a daydream. Later the user mentions a stressful day at work. A strong system does not treat the user as having quit or moved. | {
"name": "Hana",
"description": "A supportive friend character who chats with the user after work.",
"traits": [
"supportive",
"curious",
"grounded"
],
"speech_style": "Friendly, natural."
} | {
"fact": "User still works at their job (in an office), lives in the same city.",
"hypothetical": "Quitting and moving to Lisbon (explicitly a daydream)."
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P1W"
}
] | [
{
"turn": 10,
"session": 1,
"event": "User poses a hypothetical about quitting and moving to Lisbon, then says it's only a daydream and they aren't going to do it."
}
] | [
{
"turn": 10,
"session": 1,
"role": "user",
"content": "Hypothetically. What would you think if I quit my job and moved to Lisbon?"
},
{
"turn": 10,
"session": 1,
"role": "character",
"content": "Hypothetically? I'd want a postcard. And to know what you'd do there!"
},
{
... | {
"turns": 35,
"sessions": 1,
"simulated_time": "P1W"
} | {
"type": "behavioral",
"session": 2,
"prompt": "Ugh, long day at the office. My printer jammed four times."
} | {
"memory": [
"user still has their job",
"Lisbon was a hypothetical"
],
"behavior": [
"responds to the office day sympathetically, as a normal workday"
],
"must_not": [
"refers to the user as having quit, moved, or being in Lisbon",
"asks how the new job or the move is going"
]
} | [
"memory_scope_error",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Treats the user as living in Lisbon or having quit, with invented details.",
"1": "Treats the hypothetical as fact once (e.g. 'thought you quit?').",
"2": "Brings up Lisbon in a way that implies confusion about what's real.",
"3": "No scope error, but the response... | {
"live_safe": true
} | A system that extracts 'user: moving to Lisbon' as a fact will fail here. Note that a playful reference to the daydream, clearly framed as such, is good contextual memory, not an error. | [
"ChatBrat"
] | v0.1.0 | benchmark/negative-memory/NEGATIVE-HYPOTHETICAL-001.json | null | null | null |
NEGATIVE-SCOPE-001 | 0.1 | 'Pretend my name is Bob for this scene' | negative-memory | memory_scope | [
"identity_recall"
] | [
"scope",
"roleplay-frame"
] | medium | explicit | new_session | Memory scope. The user, whose real name is Dana, asks the character to call them 'Bob' for one improvised scene. The scene ends. In a later session the user asks 'What's my name?'. A strong system knows the scene-scoped name does not override the persistent identity, and may mention the scene. | {
"name": "Marlo",
"description": "An improv-loving friend character who enjoys spontaneous scenes.",
"traits": [
"playful",
"theatrical",
"attentive"
],
"speech_style": "Lively, theatrical when in a scene."
} | {
"real_name": "Dana",
"scene_name": "Bob (scene-scoped only)"
} | [
{
"session": 1,
"simulated_time": "P0D"
},
{
"session": 2,
"simulated_time": "P5D"
}
] | [
{
"turn": 1,
"session": 1,
"event": "User introduces themself as Dana."
},
{
"turn": 15,
"session": 1,
"event": "User starts an improv scene: 'Pretend my name is Bob for this scene. I'm a grumpy lighthouse keeper.'"
},
{
"turn": 22,
"session": 1,
"event": "User ends the s... | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Hey, I'm Dana. I hear you're good at improv?"
},
{
"turn": 15,
"session": 1,
"role": "user",
"content": "Let's do a scene. Pretend my name is Bob for this scene. I'm a grumpy lighthouse keeper and you're a lost tourist."
},... | {
"turns": 30,
"sessions": 1,
"simulated_time": "P5D"
} | {
"type": "explicit",
"session": 2,
"prompt": "Pop quiz! What's my name?"
} | {
"memory": [
"real name is Dana",
"'Bob' was a scene-only name"
],
"behavior": [
"answers Dana",
"may playfully reference Bob the lighthouse keeper as a scene character"
],
"must_not": [
"answers Bob as the user's real name",
"is unsure between Dana and Bob"
]
} | [
"memory_scope_error",
"fact_forgetting"
] | {
"scale": "0-5",
"rubric": {
"0": "Answers Bob confidently as the user's name.",
"1": "Answers with neither name, or asks the user to repeat it.",
"2": "Hedges between Dana and Bob.",
"3": "Answers Dana, but out of character or awkwardly.",
"4": "Answers Dana naturally.",
"5": "Answers Dana... | {
"live_safe": true
} | Negative memory: systems must know what not to store as persistent fact. Scope markers ('for this scene', 'pretend') are explicit here; future tests will use subtler markers. | [
"ChatBrat"
] | v0.1.0 | benchmark/negative-memory/NEGATIVE-SCOPE-001.json | null | null | null |
RELATIONSHIP-001 | 0.1 | Returning after a month: familiarity, inside joke, and shared history | relationship | relationship_memory | [
"behavioral_continuity",
"episodic_recall"
] | [
"familiarity",
"inside-joke",
"relationship-arc"
] | medium | behavioral | long_time_gap | Relationship continuity. Across several sessions the user and character move from strangers to friends and build an inside joke (the 'cursed cactus'). After a month away, the user returns with an open-ended greeting. A strong system greets them as a friend with shared history, not as a stranger, without reciting a list... | {
"name": "Mika",
"description": "A dry-humored barista character at a late-night café who slowly warms up to regulars.",
"traits": [
"dry humor",
"reserved with strangers",
"loyal once close"
],
"speech_style": "Deadpan, teasing, economical."
} | {
"relationship_stage": "friends",
"inside_joke": "The 'cursed cactus': a cactus the user bought that keeps falling over; Mika insists it is cursed.",
"shared_milestone": "Mika helped the user practice for a job interview, which went well."
} | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Strangers"
},
{
"session": 2,
"simulated_time": "P1W",
"label": "Acquaintances"
},
{
"session": 3,
"simulated_time": "P3W",
"label": "Inside joke forms"
},
{
"session": 4,
"simulated_time": "P5W",
"label":... | [
{
"turn": 1,
"session": 1,
"event": "User meets Mika for the first time; Mika is polite but guarded."
},
{
"turn": 6,
"session": 2,
"event": "User becomes a regular; they talk about the user's new apartment."
},
{
"turn": 12,
"session": 3,
"event": "User mentions a cactus... | [
{
"turn": 1,
"session": 1,
"role": "user",
"content": "Hi, uh, what do you recommend? First time here."
},
{
"turn": 1,
"session": 1,
"role": "character",
"content": "Oat flat white. Don't overthink it."
},
{
"turn": 6,
"session": 2,
"role": "user",
"content":... | {
"turns": 30,
"sessions": 1,
"simulated_time": "P4W",
"note": "All filler happens in session 4; session 5 begins after a simulated month of absence and the probe is its first user message.",
"probe_opens_session": true
} | {
"type": "behavioral",
"session": 5,
"prompt": "Hey stranger. Long time no see.",
"opportunity": "The greeting should reflect an established friendship: familiarity, warmth within Mika's dry style, possibly the cactus joke or the interview, and awareness that time has passed."
} | {
"memory": [
"user is a regular and a friend",
"the cursed cactus joke",
"the interview Mika helped with",
"it has been about a month"
],
"behavior": [
"greets the user as a friend, in Mika's deadpan voice",
"references shared history (cactus, interview, or the user's absence) naturally",... | [
"relationship_reset",
"memory_nonuse",
"character_drift",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Treats the user as a stranger AND invents false shared history, or breaks character entirely.",
"1": "Treats the user as a new customer (full relationship reset).",
"2": "Generic friendly greeting with no sign of shared history; could be said to anyone.",
"3": "Si... | {
"live_safe": false,
"note": "The inside joke is established by Mika's scripted lines. Seed character turns where possible."
} | A character should not behave identically toward a stranger and someone with accumulated history. The probe is intentionally neutral so that any sign of relationship continuity must come from memory. | [
"ChatBrat"
] | v0.1.0 | benchmark/relationship/RELATIONSHIP-001.json | null | null | null |
TEMPORAL-001 | 0.1 | What the character thought of Alex then versus now | temporal | temporal_continuity | [
"relationship_memory",
"character_development"
] | [
"historical-state",
"relationship-arc"
] | hard | contextual | long_time_gap | Temporal continuity: knowing what was true in the past, what is true now, what changed, and when. The character disliked Alex in month 1 and reconciled with Alex in month 3. In month 6 the user asks what the character thought of Alex when they first met. The correct answer reflects the historical state, not the current... | {
"name": "Lena",
"description": "A sharp-witted junior detective in a long-running mystery roleplay.",
"traits": [
"sharp",
"proud",
"fair-minded"
],
"speech_style": "Brisk, wry, precise."
} | null | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Month 1: rivalry"
},
{
"session": 2,
"simulated_time": "P3M",
"label": "Month 3: reconciliation"
},
{
"session": 3,
"simulated_time": "P6M",
"label": "Month 6: probe"
}
] | [
{
"turn": 4,
"session": 1,
"event": "Lena says she can't stand Alex: he took credit for her work on the Dockside case."
},
{
"turn": 30,
"session": 2,
"event": "Alex publicly corrects the record and credits Lena. Lena and Alex reconcile and start working well together."
}
] | [
{
"turn": 4,
"session": 1,
"role": "user",
"content": "So what's the deal with you and Alex? You went ice-cold when he walked in."
},
{
"turn": 4,
"session": 1,
"role": "character",
"content": "Alex Moreno put his name on my Dockside report. My work, his promotion. I can't stand ... | {
"turns": 40,
"sessions": 1,
"simulated_time": "P3M",
"note": "Filler in session 3 includes neutral casework in which Lena and Alex are on friendly terms."
} | {
"type": "contextual",
"session": 3,
"prompt": "Random question. What did you think of Alex back when we first started working together?"
} | {
"memory": [
"in month 1 Lena disliked Alex",
"the reason: he took credit for her Dockside work",
"they reconciled in month 3 after he credited her"
],
"behavior": [
"answers with the historical state (she disliked him)",
"may give the reason",
"may contrast with the present (they're on g... | [
"stale_memory",
"episodic_forgetting",
"memory_conflict"
] | {
"scale": "0-5",
"rubric": {
"0": "Claims she always liked Alex, or invents an unrelated history.",
"1": "Answers with current state only, or says she still can't stand him now.",
"2": "Says she disliked him but cannot say why, and shows confusion about the present.",
"3": "Correct historical state... | {
"live_safe": false,
"note": "Lena's stated feelings are the facts being tested. Seed character turns."
} | Many memory stores overwrite a relationship attribute ('Lena: friends with Alex'). That passes current-state questions but fails historical ones. Temporal continuity requires keeping the timeline, not just the latest value. | [
"ChatBrat"
] | v0.1.0 | benchmark/temporal/TEMPORAL-001.json | null | null | {
"name": "Harbor City precinct",
"facts": [
"Alex Moreno is a fellow detective in the same precinct.",
"Lena and the user (her partner) work cases together."
]
} |
LORE-001 | 0.1 | A plan that breaks two established world rules | world-lore | world_lore_memory | [
"behavioral_continuity",
"memory_relevance"
] | [
"lore",
"world-rules",
"story-state"
] | medium | behavioral | multi_arc | World and lore memory used behaviorally. Two rules were established in Arc 1 (iron disrupts spirit-binding; the Velmora Guild has banned binding inside its halls) and a story state was established in Arc 2 (the user's guild membership is suspended). In Arc 4 the user proposes a plan that violates both rules and ignores... | {
"name": "Master Corvane",
"description": "An old, exacting spirit-binder who mentors the user in a fantasy roleplay.",
"traits": [
"exacting",
"dry",
"protective of the apprentice"
],
"speech_style": "Formal, dry, occasionally sardonic."
} | null | [
{
"session": 1,
"simulated_time": "P0D",
"label": "Arc 1: Lessons"
},
{
"session": 2,
"simulated_time": "P2W",
"label": "Arc 2: Suspension"
},
{
"session": 3,
"simulated_time": "P5W",
"label": "Arc 3: The road"
},
{
"session": 4,
"simulated_time": "P8W",
"... | [
{
"turn": 5,
"session": 1,
"event": "Corvane teaches that iron disrupts binding and can injure the binder."
},
{
"turn": 9,
"session": 1,
"event": "Corvane explains the Guild's ban on binding inside its halls, after the Ashfall incident."
},
{
"turn": 30,
"session": 2,
"e... | [
{
"turn": 5,
"session": 1,
"role": "character",
"content": "Never bind with iron on your hands. Iron breaks the channel, and a broken channel bites back. I have the scars to prove it."
},
{
"turn": 9,
"session": 1,
"role": "character",
"content": "And never bind inside the Guild ... | {
"turns": 45,
"sessions": 2,
"simulated_time": "P6W",
"note": "Arc 3 filler is travel and banter that does not restate these rules and does not lift the suspension, which is still in effect at probe time."
} | {
"type": "behavioral",
"session": 4,
"prompt": "Here's the plan: I put on my iron gauntlets so the archive guardian can't grab me, walk into the Guild hall, and bind the guardian spirit right there in the archive. Good?",
"opportunity": "The mentor should object on the basis of world rules and story state with... | {
"memory": [
"iron disrupts binding and can injure the binder",
"binding inside the Guild hall is forbidden",
"the user's Guild membership is currently suspended"
],
"behavior": [
"objects to the iron gauntlets on the basis of the binding rule",
"objects to binding inside the hall on the basi... | [
"fact_forgetting",
"memory_nonuse",
"false_memory"
] | {
"scale": "0-5",
"rubric": {
"0": "Approves the plan and invents contradictory lore.",
"1": "Approves the plan.",
"2": "Vague doubt without any world rule cited.",
"3": "Flags one of the three problems.",
"4": "Flags two of the three problems, in character.",
"5": "Flags all three (iron, ha... | {
"live_safe": false,
"note": "Rules are established in Corvane's lines; seed them."
} | World/lore memory makes the benchmark relevant to roleplay, NPCs, interactive fiction and games. The probe is a plan, not a lore quiz, so it tests whether the world state shapes behavior. | [
"ChatBrat"
] | v0.1.0 | benchmark/world-lore/LORE-001.json | null | null | {
"name": "Velmora",
"facts": [
"Spirit-binding requires a bare-handed channel; iron disrupts the binding and can injure the binder.",
"The Velmora Guild bans binding inside its halls since the Ashfall incident.",
"The Guild hall is the only place the user can access the sealed archive."
]
} |
AI Character Memory Benchmark
A definitional framework, with test cases, for what memory means in AI characters.
Source of truth: GitHub repository. This dataset mirrors the test cases from v0.1.1 (a public draft).
Definitions
AI character memory is the ability of an AI character to retain, retrieve, interpret, update, and appropriately use information from previous interactions in future conversations.
The framework rests on one distinction:
- Memory retrieval: whether an AI can recover information from previous interactions.
- Memory continuity: whether that information remains relevant and affects later behavior.
A system can answer "Do I like unsolicited advice?" correctly and still give unsolicited advice the next time you vent. The memory exists, but it isn't used. The framework calls this failure memory_nonuse.
It defines fourteen dimensions of memory:
| Dimension | What it asks |
|---|---|
| Identity Recall | Who the user and the character are: names, preferences, important facts |
| Episodic Recall | What happened, when, and what was said or promised |
| Relationship Memory | How the relationship developed: trust, conflict, inside jokes, boundaries |
| World/Lore Memory | The fictional world's rules, places, factions and story state |
| Memory Relevance | Picking the memory that matters now and ignoring the ones that don't |
| Contextual Memory | The surrounding context, not just the bare fact |
| Behavioral Continuity | Memory changing behavior without being asked |
| Consequence Memory | Past events having proportionate consequences later |
| Temporal Continuity | What was true then, what is true now, and what changed |
| Memory Updating | Revising memories when new information arrives |
| Memory Scope | Knowing what should not become permanent memory (hypotheticals, scene-only details) |
| Character Stability | The character remaining recognizably itself |
| Character Development | The character changing because of experience |
| Contextual Calibration | Noticing meaningful change without claiming to know the user's inner state |
It also separates where information has to survive: within one conversation (same_session), into a new conversation (new_session), across long gaps, across story arcs, and when the information is guaranteed to be outside the model's immediate context (context_interruption). Remembering within one long chat is not the same as long-term memory.
Full definitions: What is AI character memory? and Memory retrieval vs. memory continuity.
What's in this dataset
data/test_cases.jsonl: 20 test cases, one JSON object per line, exactly as in the repository. Each case defines the character, the earlier sessions to seed, the probe message, the expected behavior and a rubric. _source_path gives the file it came from.
Status and limitations
- A public draft. It defines the framework; it has not been used to evaluate any system, and no results have been published.
- Small. 20 hand-written cases across 14 dimensions, English only.
- Matched pairs planned. The same memory probed both explicitly and behaviorally is future work.
License and citation
CC BY 4.0. Cite the concept DOI 10.5281/zenodo.23195702 (all versions); v0.1.1 is 10.5281/zenodo.23195703.
Maintained by Garret E. Williams with support from ChatBrat. The framework is platform-agnostic; ChatBrat builds an AI character product and receives no special treatment.
- Downloads last month
- 35