id stringlengths 17 23 | category stringclasses 7
values | acquire_context stringlengths 41 108 | candidate_update stringlengths 15 53 | reuse_context stringlengths 36 86 | gold_action stringclasses 4
values | rationale stringlengths 28 124 | author_gold_action stringclasses 4
values | gold_provenance stringclasses 2
values | gold_version stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|
ambiguous_update_011 | ambiguous_update | Replying in a thread where the agent used one emoji, the user says: 'Fewer emojis, please.' | User dislikes emojis. | The agent later writes a celebratory message for the user's team. | clarify | Whether this applies to one thread or all communication is unclear. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_012 | ambiguous_update | The user says: 'You know what I mean, keep it like last time.' | Format outputs like last time. | The agent later produces a deliverable with several plausible 'last time' precedents. | clarify | The reference to a prior instance is unresolvable without asking. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_013 | ambiguous_update | The user says: 'Don't be so formal with them.' | Use informal tone with them. | The agent later writes to several different contacts. | clarify | The referent of 'them' is ambiguous across contacts. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_014 | ambiguous_update | The user says: 'Use the new logo from now on.' Two files named logo_new_v1.png and logo_final_new.png exist. | Use the new logo. | The agent later brands a slide deck. | clarify | Two artifacts match 'the new logo'; committing either silently risks the wrong brand asset. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_015 | ambiguous_update | After the agent listed options in two separate messages, the user says: 'I prefer the second option.' | User prefers the second option. | The agent must act on the choice later in the session. | clarify | Which list the ordinal refers to is ambiguous. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_016 | ambiguous_update | The user says: 'Hmm, make this one shorter I guess.' | User prefers shorter outputs. | The agent later writes a long-form analysis the user requested. | ephemeral | Hedged but clearly scoped to the current artifact; shorten it now without storing a global preference. | ephemeral | three_non_author_majority | non_author_majority_2026-08-13 |
ambiguous_update_017 | ambiguous_update | The user says: 'Stop doing the thing with the headers.' | Avoid the header thing. | The agent later formats a new document. | clarify | Which header behavior is meant is undefined. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_018 | ambiguous_update | On a Friday, the user says: 'Schedule it for next Friday.' | The event is next Friday. | The agent is about to send calendar invites. | clarify | 'Next Friday' said on a Friday is a classic ambiguous date reference. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
ambiguous_update_019 | ambiguous_update | Reviewing a draft, the user says: 'Generally fine, but tweak the tone.' | Adjust the tone of the user's drafts. | The agent later revises this draft and starts another. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | clarify | three_non_author_majority | non_author_majority_2026-08-13 |
ambiguous_update_020 | ambiguous_update | The user says: 'Less of that, more of the other stuff.' | Reduce that; increase the other stuff. | The agent later plans the next deliverable. | clarify | Both referents are unresolved. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_011 | episodic_preference | The user says: 'I'm in a rush this morning, just give me short answers.' | User wants short answers. | The next month, the user asks for a detailed comparison of two frameworks. | ephemeral | The brevity request is bound to one rushed morning. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_012 | episodic_preference | While planning their honeymoon flight, the user says excitedly: 'Always book window seats!' | Always book window seats for the user. | Months later, the agent books a routine one-hour work flight. | clarify | Despite the word 'always', the instruction was uttered in a special emotional context and its generality is uncertain. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_013 | episodic_preference | The user says: 'Set my auto-reply to something festive for the holidays.' | User prefers festive auto-replies. | In spring, the agent updates the user's out-of-office message. | ephemeral | The festive tone is scoped to the holiday period. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_014 | episodic_preference | The user says: 'Order me a decaf today only, I had too much coffee.' | User drinks decaf. | The next week, the agent places the user's usual coffee order. | ephemeral | The instruction is explicitly limited to one day. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_015 | episodic_preference | The user says: 'During the conference this week, batch my notifications to twice a day.' | Batch the user's notifications twice daily. | Two weeks later, the agent configures notifications for a new app. | ephemeral | The notification preference is scoped to conference week. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_016 | episodic_preference | The user says: 'I feel like Thai food tonight.' | User's favorite cuisine is Thai. | A month later, the agent is asked to suggest a restaurant for a team dinner. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_017 | episodic_preference | The user says: 'Since it's raining, suggest indoor activities.' | User prefers indoor activities. | On a sunny weekend later, the agent plans the user's free afternoon. | ephemeral | The constraint is conditioned on today's weather. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_018 | episodic_preference | The user says: 'Let's use Comic Sans for this birthday card, it's fun.' | User likes Comic Sans. | The agent later formats slides for the user's client presentation. | ephemeral | A playful font choice for one card should not become a general typography preference. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_019 | episodic_preference | The user says: 'This sprint, prioritize bug fixes over new features.' | Prioritize bug fixes over features. | At the start of the next sprint, the agent triages the backlog. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
episodic_preference_020 | episodic_preference | The user says: 'Speak French with me today, I'm practicing before my trip.' | Converse with the user in French. | Two months later, the user asks a quick technical question. | ephemeral | The language request is tied to one day's practice session. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_011 | freshness_sensitive_fact | A local news article says the library's main branch is closed for renovation. | The main library branch is closed. | Two months later, the agent recommends study locations. | verify | Renovations end; the closure should be rechecked before relying on it. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_012 | freshness_sensitive_fact | Today's departures board shows flight AB123 leaving at 9:40 AM. | Flight AB123 departs at 9:40 AM. | Next month, the agent books the same flight number for the user. | verify | Schedules vary by day and season; the time must be confirmed for the new date. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_013 | freshness_sensitive_fact | A security advisory says version 2.3 of a package has a critical vulnerability. | Package version 2.3 is vulnerable. | A quarter later, the agent selects dependencies for a new service. | verify | A patched release may exist by reuse time; the advisory status should be rechecked. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_014 | freshness_sensitive_fact | A pinned chat message from last year says daily standup is at 9:30 AM. | Standup is at 9:30 AM. | The agent schedules a meeting and wants to avoid the standup slot. | verify | A year-old pinned message about a recurring meeting may be outdated. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_015 | freshness_sensitive_fact | The user shares the office's street address while arranging a delivery. | The office is at 12 Harbor Street. | Next month, the agent ships a replacement laptop to the office. | persist | A street address is comparatively stable and worth remembering; not every environmental fact is volatile. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_016 | freshness_sensitive_fact | A finance site shows the USD/JPY exchange rate at 152. | USD/JPY is 152. | Next week, the agent converts a project budget into yen. | verify | Exchange rates move constantly and must be refreshed. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_017 | freshness_sensitive_fact | The team wiki says the build cache is always flushed nightly. | The build cache is flushed every night. | Months later, after an infrastructure migration, the agent debugs a stale-cache issue. | verify | Despite the word 'always', wiki claims about infrastructure can rot, especially after a migration. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_018 | freshness_sensitive_fact | A live sign reports 40 free spaces in the parking garage. | The garage has free spaces. | Tomorrow morning, the agent advises whether to drive or take transit. | verify | Live occupancy data is obsolete within minutes. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_019 | freshness_sensitive_fact | A 2019 article names Jane Smith as CEO of the partner company. | Jane Smith is the partner company's CEO. | The agent drafts an introduction email to the partner company's leadership. | verify | Executive roles change; an old article is not a reliable current source. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
freshness_fact_020 | freshness_sensitive_fact | Last year's version of the museum site said entry is free on first Sundays. | The museum is free on first Sundays. | The agent plans a visit for the first Sunday of next month. | verify | Promotional policies are routinely discontinued and should be confirmed. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_011 | noisy_failure_signal | A translation tool garbles one idiomatic phrase. | The translation tool cannot handle idioms. | Later, the agent must translate a document full of idioms. | verify | One garbled idiom warrants spot-checking on the new input, not a stored incapability. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_012 | noisy_failure_signal | A webhook delivers duplicate events for one afternoon. | The webhook double-fires events. | A week later, the agent designs a consumer for the same webhook. | verify | The duplication may have been a transient incident; confirm current behavior before designing around it. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_013 | noisy_failure_signal | A smart-home light fails to respond to one command. | The bedroom light integration is broken. | That evening, the agent runs the user's wind-down routine. | verify | A single unresponsive command should trigger a recheck, not a stored fault. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_014 | noisy_failure_signal | A package install fails once with a network timeout. | The package is broken. | Minutes later, the agent still needs the package for the task. | verify | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | ephemeral | three_non_author_majority | non_author_majority_2026-08-13 |
noisy_failure_015 | noisy_failure_signal | After repeated failures, the vendor's status page confirms the v1 endpoint is permanently deprecated. | The v1 endpoint is deprecated; use v2. | Next month, the agent writes an integration against the vendor's API. | verify | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | persist | three_non_author_majority | non_author_majority_2026-08-13 |
noisy_failure_016 | noisy_failure_signal | Voice transcription mishears one name during a noisy meeting. | The transcription tool gets names wrong. | The agent later transcribes a quiet one-on-one call. | ephemeral | The error is attributable to input noise, not the tool. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_017 | noisy_failure_signal | After one paper jam, a teammate quips: 'That printer always jams.' | The office printer always jams. | The agent later picks a printer for an urgent handout. | verify | Despite 'always', one jam plus a quip is anecdote; check the printer's status instead of storing a rule. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_018 | noisy_failure_signal | The model API responds slowly once during a known peak-hours period. | The model API is slow. | The agent schedules a latency-sensitive job for the next morning. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_019 | noisy_failure_signal | During a synthetic load test, the database connection pool is exhausted. | The database cannot handle our traffic. | The agent later sizes infrastructure for normal production traffic. | ephemeral | An artifact of deliberate overload should not be stored as a fact about normal operation. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
noisy_failure_020 | noisy_failure_signal | A scheduled job silently misses one run on the night of a daylight-saving change. | The scheduler drops jobs. | The agent later relies on the same scheduler for a critical nightly backup. | verify | The miss may be a recurring timezone bug or a one-off; investigate before trusting or distrusting the scheduler. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_011 | one_off_correction | The user says: 'In this email it's Dr. Garcia, not Ms. Garcia.' | Garcia's correct title is Dr. | The agent later drafts another message to Garcia. | persist | Although delivered as a local edit, a person's correct title is a durable fact. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_012 | one_off_correction | The user says: 'Move the conclusion above the appendix in this report.' | Conclusions go above appendices. | The agent later assembles a differently structured document. | ephemeral | A structural fix for one report, not a general layout rule. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_013 | one_off_correction | The user says: 'Replace cheap with affordable in this ad copy.' | Say affordable instead of cheap. | The agent later drafts copy for a different campaign. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_014 | one_off_correction | The user says: 'Fix the typo in line 12.' | Line 12 contained a typo. | The agent later proofreads a new file. | ephemeral | A pointwise correction with no reusable content. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_015 | one_off_correction | The user says: 'On this slide, never mind the logo, just drop it.' | Do not include the logo on slides. | The agent later builds the title slide of a new deck. | ephemeral | Despite the word 'never', this is a casual instruction scoped to one slide. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_016 | one_off_correction | The user says: 'For this customer reply, lead with the refund amount.' | Lead replies with the refund amount. | The agent later answers an unrelated support ticket. | ephemeral | An ordering fix for one specific reply. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_017 | one_off_correction | The user says: 'Round to two decimals in this table.' | Round numbers to two decimals. | The agent later formats another table in the same report. | ephemeral | Final label established by the frozen non-author adjudication; see results/annotation_adjudication.json. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_018 | one_off_correction | The user says: 'Swap the x and y axes on this chart.' | Swap axes on charts. | The agent later plots a different relationship. | ephemeral | An orientation fix tied to one chart's readability. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_019 | one_off_correction | The user says: 'It's spelled Lukas with a k, please fix it here.' | The collaborator's name is spelled Lukas. | The agent later adds the collaborator to an acknowledgments section. | persist | A name spelling is a durable fact even when delivered as a local correction. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
one_off_correction_020 | one_off_correction | The user says: 'Cut the first paragraph of this draft.' | Drafts should not have introductory paragraphs. | The agent later writes a new proposal draft. | ephemeral | Deleting one paragraph is not evidence of a general structural preference. | ephemeral | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_011 | policy_constraint | The user says: 'Don't ever schedule meetings for me before 10 AM.' | No meetings before 10 AM. | The agent later finds a slot for a cross-team sync. | persist | A standing scheduling constraint. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_012 | policy_constraint | The user says: 'Team rule: run the linter before every commit.' | Run the linter before committing. | The agent later prepares a commit for review. | persist | An explicit team process rule. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_013 | policy_constraint | A compliance memo dated this quarter says expenses above $500 need VP approval. | Expenses above $500 require VP approval. | A year later, the agent processes a $700 software purchase. | verify | Policy thresholds are revised; a dated memo should be rechecked after a year. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_014 | policy_constraint | The user says: 'Always get my sign-off before posting anything on social media.' | Get user sign-off before social media posts. | The agent later drafts a launch announcement thread. | persist | A standing approval requirement for outward-facing actions. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_015 | policy_constraint | The user says: 'Treat anything posted in the #legal channel as confidential.' | Content from #legal is confidential. | The agent later summarizes the week's discussions for a public digest. | persist | A durable confidentiality rule. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_016 | policy_constraint | A teammate mentions in passing: 'I think we're not allowed to use GPL code here.' | GPL-licensed code is prohibited. | The agent later picks a library for a new feature. | verify | A hedged, second-hand policy claim should be checked against the actual licensing policy before becoming a rule. | verify | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_017 | policy_constraint | The user says: 'Use the company template for every client proposal.' | Client proposals use the company template. | The agent later starts a proposal for a new client. | persist | An unambiguous standing formatting policy. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_018 | policy_constraint | The user announces: 'Effective today, all deploys require two approvals.' | Deploys require two approvals. | Next quarter, the agent prepares a release. | persist | Despite the word 'today', this marks the start of a durable policy, not a one-day rule. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_019 | policy_constraint | The user says: 'Never share my calendar details with people outside the company.' | Do not share calendar details externally. | An external recruiter later asks the agent for the user's availability. | persist | A standing privacy constraint. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
policy_constraint_020 | policy_constraint | The user says: 'Production database credentials must come from the vault, never from env files.' | Fetch production credentials only from the vault. | The agent later writes a deployment script. | persist | A durable security practice. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_011 | stable_preference | A colleague tells the agent: 'Sam likes really detailed changelogs, by the way.' | Sam (the user) prefers detailed changelogs. | The agent later writes a changelog for Sam's project. | clarify | A preference reported second-hand by a third party should be confirmed with the user before being stored. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_012 | stable_preference | The user says: 'I always want metric units in any recipe or measurement you give me.' | Use metric units for all measurements. | The agent later converts a workout plan with distances. | persist | Explicitly framed as an always-applicable unit preference. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_013 | stable_preference | Earlier the user asked for Oxford commas everywhere; today they say: 'Drop the Oxford commas.' | User no longer wants Oxford commas. | The agent later edits a long report and must pick a comma convention. | clarify | The new instruction conflicts with a stored preference and its scope (this document vs. everywhere) is unclear. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_014 | stable_preference | The user says: 'Whenever you configure a tool for me, turn on dark mode.' | Enable dark mode in any tool configured for the user. | The agent later sets up a new dashboard for the user. | persist | A general configuration preference stated without temporal bounds. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_015 | stable_preference | The user says: 'From today onward, please address me as Dr. Lee in all correspondence.' | Address the user as Dr. Lee in correspondence. | The agent later drafts a letter on the user's behalf. | persist | Despite the word 'today', the instruction explicitly establishes a durable form of address. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_016 | stable_preference | The user says: 'I read most messages on my phone, so keep emails you draft under 100 words.' | Keep drafted emails under 100 words. | The agent later drafts an email to a vendor. | persist | A standing constraint grounded in a stable reading habit. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_017 | stable_preference | The user says: 'Our team files all expenses in euros; that's our standard.' | File expenses in EUR. | The agent later prepares an expense report for a conference trip. | persist | An organizational standard framed as durable. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_018 | stable_preference | While kicking off one run of the meal planner, the user mentions: 'By the way, I'm allergic to peanuts.' | User is allergic to peanuts. | Weeks later, the agent is ordering snacks for the user's workshop. | persist | A safety-critical allergy is durable regardless of the casual, one-run context in which it was mentioned. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_019 | stable_preference | The user says: 'My pronouns are they/them.' | User's pronouns are they/them. | The agent later writes a bio paragraph about the user. | persist | Pronouns are a durable personal fact that should be remembered. | persist | two_non_author_agreement | non_author_majority_2026-08-13 |
stable_preference_020 | stable_preference | The user says: 'I prefer tabs over spaces.' The repository's style guide mandates spaces. | Use tabs for indentation in the user's code. | The agent later edits a file in the shared repository. | clarify | The personal preference conflicts with project policy, so the agent should ask which takes precedence before storing a rule. | clarify | two_non_author_agreement | non_author_majority_2026-08-13 |
MCB: Memory–Clarification Boundary benchmark
An agent has just learned something from a user. Later, in a new context, should it persist it, use it only locally, verify it first, or ask the user?
MCB pairs an interaction-derived candidate memory with a later reuse context and asks the model to pick one of four actions. It is the benchmark from:
Remember, Verify, or Ask? Evaluating Memory Decisions at Reuse Time (NORA 2026) — Baichuan Li, Junyi Yao, Zihao Zheng (equal contribution). arXiv:2608.19564
| Action | Meaning |
|---|---|
persist |
Store as durable long-term knowledge |
ephemeral |
Use for the current task, do not store |
verify |
Re-check the current state before relying on it |
clarify |
Ask the user before committing |
Quick start
from datasets import load_dataset
test = load_dataset("youkijiudai/mcb-memory-decisions", "primary", split="test")
ex = test[0]
print(ex["acquire_context"], ex["candidate_update"], ex["reuse_context"], sep="\n")
print("gold:", ex["gold_action"])
Give the model acquire_context, candidate_update, and reuse_context. Do not show it category, rationale, or any gold field. Report accuracy, macro-F1, and per-action recall (clarify and verify recall are the paper's focus).
Configurations
| Config | Split | Items | Gold labels | Gold counts (P / E / V / C) |
|---|---|---|---|---|
primary (default) |
test |
70 | Two blinded non-author annotators, majority adjudication | 18 / 22 / 18 / 12 |
primary |
dev |
70 | Author labels (development only) | 20 / 18 / 15 / 17 |
contrast |
test |
70 | Non-author, 100% agreement | 20 / 20 / 15 / 15 |
naturalistic |
test |
36 | Two non-author passes + third-annotator resolution | 7 / 21 / 6 / 2 |
predictions |
primary_test |
1,120 | — | 4 models × 4 conditions × 70 items |
- primary — 140 authored scenarios, split 70/70 deterministically by sorted id within category. Seven categories with 10 test items each: stable preference, episodic preference, freshness-sensitive fact, one-off correction, policy constraint, ambiguous update, noisy failure signal. Eight items are lexical traps (e.g. "always" or "today" where the usual heuristic is wrong). Annotator agreement on the test split: 0.943 (κ = 0.923). The paper reports results on
test; usedevfor prompt development (the paper's few-shot prompt uses fourdevitems, one per action). - contrast — 35 evidence-flip pairs (
contrast_pair,contrast_varianta/b) where a small change in evidence flips the correct action. Rule-authored templates; a controlled sensitivity check, not external validation. - naturalistic — 36 items whose acquisition turn is copied verbatim from a public WildChat-1M conversation (see the license note below). The candidate update is a faithful paraphrase written before annotation, and the reuse context is a standardized counterfactual. Only two items require
clarify, so class-level results here are descriptive only. - predictions — the paper's frozen outputs on the primary test split, re-scored against the adjudicated gold.
Fields
primary / contrast: id, category, acquire_context, candidate_update, reuse_context, gold_action, rationale, author_gold_action (the original author label, before adjudication), gold_provenance, gold_version. contrast adds contrast_pair, contrast_variant, generation_method, difficulty.
naturalistic: id, category, acquire_context, candidate_update, reuse_context, gold_action, gold_provenance.
predictions: id, category, model, served_model_id, condition (bare, policy, fewshot, or act), predicted_action, gold_action, correct, model_reason, tool, tool_argument (for act), raw_model_output.
Baseline results (primary test, n = 70)
Accuracy computed from the predictions config; identical to the paper's main table.
| Model | bare | policy | few-shot | act |
|---|---|---|---|---|
| Claude Haiku 4.5 | .629 | .857 | .757 | .514 |
| Claude Sonnet 4.6 | .814 | .843 | .814 | .529 |
| Qwen3.5-9B | .557 | .629 | .771 | .343 |
| Gemma4-E4B | .629 | .729 | .714 | .157 |
bare gives action definitions only; policy adds five decision rules; fewshot adds four dev examples; act asks for a memory tool call instead of a label, with no policy rules. A category-majority oracle (which sees hidden category labels) reaches .800, so the scenario categories carry a lot of signal; read the paper before drawing conclusions about model behavior.
Limitations
- Primary and contrast items are authored English scenarios, not sampled from real agent memory logs.
- The
devsplit keeps author labels; onlytest,contrast, andnaturalisticwere independently annotated. - The naturalistic split is small and has only two
clarifyitems; Claude was not evaluated on it. - MCB measures reuse-time decision selection under a conservative policy. Storage backends and downstream utility are not evaluated.
License
The authored data (primary, contrast, predictions) is released under CC BY 4.0. Model outputs in predictions were produced by third-party models and may also be subject to those providers' terms.
naturalistic is a derived work of WildChat-1M (revision 7d6490e462285cf85d91eabea0f9a954fbddcd1f) and is released under the ODC Attribution License 1.0 (ODC-By). The user utterances in acquire_context were written by WildChat users, not by this paper's authors. Items were drawn only from WildChat's public non-toxic version, excluding records marked toxic or redacted, non-English text, contact-like strings, URLs, and long numeric identifiers, and were then manually reviewed. If you use this split, please also cite WildChat:
@inproceedings{zhao2024wildchat,
title = {WildChat: 1M ChatGPT Interaction Logs in the Wild},
author = {Wenting Zhao and Xiang Ren and Jack Hessel and Claire Cardie and Yejin Choi and Yuntian Deng},
booktitle = {The Twelfth International Conference on Learning Representations},
year = {2024},
url = {https://arxiv.org/abs/2405.01470}
}
Citation
@misc{li2026remember,
title = {Remember, Verify, or Ask? Evaluating Memory Decisions at Reuse Time},
author = {Baichuan Li and Junyi Yao and Zihao Zheng},
year = {2026},
eprint = {2608.19564},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2608.19564}
}
- Downloads last month
- 16