Dataset Viewer
Auto-converted to Parquet Duplicate
system
stringclasses
6 values
instruction
stringlengths
383
55.4k
output
stringlengths
109
12.5k
metadata
unknown
You answer a question about a user from evidence collected from a memory of the user's earlier conversations with an assistant: the Core memory and the retrieved memory entries listed below. Each entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question....
Core memory: 25yo software engineer & senior UX designer in Milan, Italy. Fluent Italian. Values family, authenticity, independence, privacy rights. Recently lost grandmother, Rachel's best friend's mom, John's dad. Close with best friend Rachel, cousin Rachel, sister. Strained relationship with controlling father (bus...
<tool_call> {"name":"answer","arguments":{"reasoning":"The question asks for a specific commitment Patagonia made in 2019 regarding tax savings from a conversation about balancing stakeholder needs. The Core memory only says the user admires Patagonia and values stakeholder balance, transparency, and employee empowerme...
{ "row_id": "v5:8:run18-hybrid-evidence-drop:gpt4_1d80365e__64a6e2bdf1:2:0", "row_role": "answer", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": 684, "source": "callback-decision", "error_type": 8, "run": "run18-hybrid-evidence-drop", "trajectory": "runs/run18-hybrid-evidenc...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: Marketing Specialist at XYZ Corporation (2+ years), reports to Rachel. Python, Pandas, Matplotlib, Seaborn. Planning data analytics certification. Construction lawyer (NSW): building regulations, contracts, homeowner protection, educational content. Full-stack developer: React, JavaScript, Node.js, Mongo...
<tool_call> {"name":"search_memory","arguments":{"query":"automated formula prompts UI process product characteristics audience selection","top_k":8,"banks":["episodic","semantic","procedural"]}} </tool_call>
{ "row_id": "v5:v3:train:6ade9755:synthetic:membuilder_flat_after_v1:sft:6ade9755:S35:72df23d381d94d99a65da6fef7c86bbb:1:retrieval:1", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "source": "v3", "error_type": null, "loss_from_char": null, "v3_row_id": "train:6ade9755:...
You answer a question about a user from evidence collected from a memory of the user's earlier conversations with an assistant: the Core memory and the retrieved memory entries listed below. Each entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question....
Core memory: User Profile: - English instructor at Texas community college (Distance Ed/Educational Technology); detail-oriented, methodical, data-driven - Bilingual English/Spanish; learning Mandarin - Lives on Indiana farm; moved to UK for Edinburgh study abroad (visa approved May 27, 2023) Family: Brother graduated...
<tool_call> {"name":"answer","arguments":{"reasoning":"The retrieved entries dated 27 May 2023 directly address applicants from India: m000251 and m000243 say Japanese Encephalitis vaccine may be recommended if the person is from rural India or spent significant time in rural areas, Hepatitis B vaccination should be up...
{ "row_id": "v5:6:run29ds-stale-core:d01c6aa8__96f8754925:3:0", "row_role": "answer", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": null, "source": "callback-redo", "error_type": 6, "run": "run29ds-stale-core", "trajectory": "runs/run29ds-stale-core/d01c6aa8__96f8754925.json...
Manage MemBuilder episodic memory. You are the Episodic Memory Manager. Manage time-ordered event memories. Episodic Memory stores time-ordered, event-based information from interactionsβ€”essentially, the "diary" of user events. Each episodic memory MUST include: (a) summary: Short textual summary of the event (concis...
{"bank": "episodic", "conversation_timestamp": "2023/05/21 (Sun) 11:03", "messages": "[REAL_USER]: I'm looking for some advice on interior design. I just rearranged my living room furniture and added a new coffee table, but I'm not sure what kind of decorations would go well with it.\n[AI_ASSISTANT]: Exciting! Rearrang...
<tool_call> {"name":"manage_memory","arguments":{"bank":"episodic","operations":[{"action":"UPDATE","old_memory":"[EPISODIC] 2023-05-19: Real User organized closet after months of procrastination | Details: Last Sunday from conversation date of 2023-05-21 11:03 (calculated as 2023-05-19), Real User finally organized th...
{ "row_id": "v5:2:run29ds-value-drop:4bc144e2__d9e2e90a45:2:1", "row_role": "episodic", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": null, "source": "callback-redo", "error_type": 2, "run": "run29ds-value-drop", "trajectory": "runs/run29ds-value-drop/4bc144e2__d9e2e90a45.js...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: Tzu-li Liu. Neuroscience/biomedical engineering researcher (brain-machine interfaces, neuromodulation). Software sales using MEDDPICC. Writing thesis on social media marketing impact. Uses Obsidian, detail-oriented, data-driven. Creative: Sculptor (clay/metal/wood), guitarist learning violin (Suzuki metho...
<tool_call> {"name":"finish","arguments":{"ranked_ids":["m000130","m000452","m000129","m000131","m000132","m000133"],"revisit_agent":"none","feedback":""}} </tool_call>
{ "row_id": "v5:9:run29ds-distractor-rank:gpt4_15e38248__667f562f4c:1", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": null, "source": "hard-negative", "error_type": 9, "run": "run29ds-distractor-rank", "trajectory": "runs/run29ds-distractor-rank/gpt4...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: Creative professional near Denver. Married, family with countryside lake house. Interests: AI image generation (dystopian/anime), BJJ white belt, blockchain/crypto (Flow), game dev with NFT/Web3. Philosophy: Alan Watts, Hindu mythology (Kali), Buddhism. Creates spirituality manuscripts/YouTube scripts. M...
<tool_call> {"name":"finish","arguments":{"ranked_ids":["m000554","m000553","m000372"],"revisit_agent":"none","feedback":""}} </tool_call>
{ "row_id": "v5:2:run29ds-value-drop:1c0ddc50__99103baaa3:3:1", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": null, "source": "callback-redo", "error_type": 2, "run": "run29ds-value-drop", "trajectory": "runs/run29ds-value-drop/1c0ddc50__99103baaa3.j...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: **CRITICAL: This is an asylum seeker in Canada, NOT Stephen Donohoe** Asylum seeker in Canada. Received asylum approval February 20, 2023. Location in Canada unspecified. Language Learning: Attending LINC program since November 2020, currently level 5. Teacher: Mrs. Patel. Made friends including someone ...
<tool_call> {"name":"search_memory","arguments":{"query":"bathroom deep cleaning order steps Sunday","top_k":5,"banks":["episodic","procedural"]}} </tool_call>
{ "row_id": "v5:v3:train:06db6396:synthetic:membuilder_flat_after_v1:sft:06db6396:S49:72df23d381d94d99a65da6fef7c86bbb:1:retrieval:1", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "source": "v3", "error_type": null, "loss_from_char": null, "v3_row_id": "train:06db6396:...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: Emily Rose Thompson (post-divorce). Founder DesignSpark (Jan 15), creates Brand Style Guide Templates. Builds remote teams prioritizing emotional intelligence. Thesis: social media's mental health impact. Projects: "Black Community Impact" gamification, Win'it e-commerce. Designs APIs/databases. Developin...
<tool_call> {"name":"search_memory","arguments":{"query":"motorcycle upgrades exhaust suspension tuning bike","top_k":10,"banks":["episodic","semantic","procedural"]}} </tool_call>
{ "row_id": "v5:15:run29ds-unanswerable:1e043500__f0e65f49c2:3:2", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "loss_from_char": null, "source": "callback-redo", "error_type": 15, "run": "run29ds-unanswerable", "trajectory": "runs/run29ds-unanswerable/1e043500__f0e6...
You search a memory built from a user's earlier conversations with an assistant, to collect the evidence needed to answer a question about the user. Each memory entry has a memory_id, the dates of the sessions it came from, and its content. The question date is "now" for the question. Search rules: - Core below is con...
Core memory: Bilingual (English/Turkish). Developing AI fashion tech: Fashionoscope (B2B) and Stiloni (B2C mobile app). Product management/MVP expertise. Background: NSW residential construction law/consulting, HIA contracts, marketing/copywriting. New Haven, CT homeowner (Jan 2023). Vegan, budget-conscious, environme...
<tool_call> {"name":"search_memory","arguments":{"query":"Convict Lake Campground elevation","top_k":5,"banks":["episodic","semantic"]}} </tool_call>
{ "row_id": "v5:v3:train:gpt4_fa19884d:synthetic:membuilder_flat_after_v1:sft:gpt4_fa19884d:S37:72df23d381d94d99a65da6fef7c86bbb:3:retrieval:1", "row_role": "retrieval", "format_protocol": "memory_banks_tool_call_v5_callback", "source": "v3", "error_type": null, "loss_from_char": null, "v3_row_id": "train...
"Manage MemBuilder episodic memory.\nYou are the Episodic Memory Manager. Manage time-ordered event (...TRUNCATED)
"{\"bank\": \"episodic\", \"conversation_timestamp\": \"2023/05/28 (Sun) 04:11\", \"messages\": \"[R(...TRUNCATED)
"<tool_call>\n{\"name\":\"manage_memory\",\"arguments\":{\"bank\":\"episodic\",\"operations\":[{\"ac(...TRUNCATED)
{"row_id":"v5:4:run26-memory-review-checklist:3ba21379__23__episodic:negative","row_role":"episodic"(...TRUNCATED)
End of preview. Expand in Data Studio

MemLoop SFT v5: proactive callback (feedback) data

Supervised fine-tuning data that teaches the three agents of a memory system (memory, retrieval, answer) to notice a problem on their own, write a specific feedback, and redo the work after receiving feedback ("callbacks"). All rows are built on the LongMemEval SFT split; no dev or test data is included.

Format ("prompt I")

Every agent answers with exactly one tool call. Its arguments end with two callback fields:

  • revisit_agent: none, the agent itself, or an upstream agent (memory ← retrieval ← answer).
  • feedback: empty for none, otherwise a short, specific description of what is wrong or missing.

An agent asked to redo its work sees Feedback from <agent>: in its input, and every input ends with Callback budget: N (1 callback per agent, at most 3 per trajectory). Memory inputs are JSON; the feedback is a feedback field there. The answer agent's prompt asks for written-month dates in both reasoning and answer (ISO dates in those two fields were rewritten).

Files

Path Content
sft/train.jsonl 40,320 training rows: system, instruction, output, metadata
sft/tokenized/ The same rows tokenized for Qwen3-4B-Instruct-2507 (qwen3_nothink template, LLaMA-Factory tokenized_path), with loss masking applied (labels == -100)
sft/reserve.jsonl Spare v3 rows used only to replace rows dropped for length (none were dropped)
sft/report.json Per-type counts, row counts by role and source, tokenization checks
trajectories/<run>.jsonl.gz Raw callback trajectories, one file per generation run; each line is one trajectory (steps, per-call calls with inputs and responses, construction = what was deliberately broken)
question-pools/*.jsonl.gz Newly written questions (synthetic date / counting / listing / granularity questions; preference questions), each verified by an independent second call
quality/sft-v5.jsonl Four-check quality review of every callback trajectory
manifest.json Per run: error type, models, number of trajectories, skipped constructions, sha256

Row metadata

  • row_role: core, episodic, semantic, procedural (memory banks), retrieval, answer.
  • source: callback-decision (the call that issued a callback), callback-redo (calls after receiving feedback), hard-negative (a trap was present, no callback, final answer correct), v3 (earlier SFT rows converted to this format with revisit_agent: none).
  • error_type: 1–16 (below); run and trajectory point into trajectories/.
  • loss_from_char: for callback-decision rows, loss starts at this character of output (the callback fields). The flawed first output stays visible but is not trained. Other rows are trained on the full output.

Composition

Role Rows Share
retrieval 20,644 51.2%
answer 9,636 23.9%
memory (core / episodic / semantic / procedural) 10,040 (2,970 / 2,372 / 2,291 / 2,407) 24.9%

Sources: callback-decision 4,384, callback-redo 12,313, hard-negative 4,101, v3 19,522. Trained tokens: 7,113,495; longest row 14,568 tokens.

16 error types (callback trajectories used, target 300 each)

# Error type Stage Passed rules Passed review Used Hard negatives
1 Memory gap (a session's fact never written) memory 810 710 300 115
2 Value dropped (vague entry, value missing) memory 404 295 295 150
3 Wrong fact in memory (incl. two conflicting values) memory 548 511 300 150
4 Memory self-review with a checklist memory 418 (rule 4) 280 150
5 Wording mismatch (question words don't retrieve the entry) memory/retrieval 233 151 142 150
6 Stale Core (Core holds an older value) retrieval/answer 1,011 885 300 150
7 Retrieval gap (search stopped too early) retrieval 403 353 300 150
8 Evidence dropped retrieval 758 719 300 103
9 Distractor ranked first answer 906 772 300 150
10 Distractor vs. Core answer 788 672 300 150
11 Wrong granularity (city vs. state) answer 26 19 19 150
12 Date arithmetic / temporal reasoning answer 440 379 300 150
13 Counting answer 226 171 132 150
14 Listing answer 501 442 300 150
15 Unanswerable (false premise) answer 404 259 259 9
16 Preference / advice answer 552 479 300 150

Types 5, 11 and 13 ran out of source questions and stay below 300.

How it was generated

  • Memory: teacher memory written by Claude Sonnet 4.5 over the LongMemEval SFT split (50 histories), then deliberately broken per error type (drop or change a value, add a distractor entry, move the true fact to Core, add an older value to Core, …) or left intact with a newly written question.
  • Agents: DeepSeek-V4.1-Flash and GLM-5.3-Flash (self-hosted) as retrieval and answer agents. DeepSeek-V4.1-Flash also writes and redoes memory and judges correctness. When an agent fails without calling back, the teacher writes the feedback the agent should have written from the reference answer ("hindsight feedback"), restricted to what that agent could see; the agent then redoes the step. Some types resample the first answer or retrieval at temperature 1.0 to find a natural mistake.
  • Filtering: rule-based filters (callback executed, clean format, final answer correct, memory rewrite verified, feedback does not leak the answer or talk about format, dedup), then a DeepSeek review of every callback trajectory on four checks: cause correct, specific, uses no hidden information, redo follows the feedback. Pass rate on all four checks: 84% for DeepSeek-agent trajectories, 82% for GLM-agent trajectories.

Source and license

Conversations and questions come from LongMemEval (xiaowu0162/longmemeval-cleaned, MIT). This dataset is released under the MIT license.

Downloads last month
-