mempabench/mempabench-traces

MemPABench traces

The conversations, tool calls and scores of the runs reported in the MemPABench paper, https://arxiv.org/abs/2610.30485. The evaluation code is at https://github.com/mempabench/mempabench, and the four simulated users it runs are the dataset https://huggingface.co/datasets/mempabench/mempabench-personas.

The runs

288 probe rounds: 4 users (user_a to user_d) x 4 PA models x 6 conditions x 3 rounds. The PA models are openai/gpt-5.6-sol, qwen/qwen3.8-27b, tencent/hy4-preview and deepseek/deepseek-v4-flash (OpenRouter ids).

• Memory conditions: nanobot_file_memory (File memory), simple_rag (Simple RAG), mem0 (Mem0) and memos (MemOS). Each memory setting (user x PA model x memory condition) first ran a session phase: the user's sessions in timeline order, 94 everyday sessions (acc_NNN) and 5 preference-change events (evolv_NN), with the memory carried from session to session. Its three probe rounds start from checkpoints of that session phase.
• Reference conditions: no_memory (No Memory: the probe runs without memory) and oracle_preference (Oracle: no memory except the probe's true preference setting, given where retrieved memory goes). They have no session phase of their own: their probes start from checkpoints of the run named by source_run_id, the user's DeepSeek-V4-Flash x File memory run.

A probe round runs 33 probes: pre_01 to pre_05 before preference changes and fin_01 to fin_28 at the end of the timeline. The offline judge (google/gemini-3.8-flash) scores each probe. The paper's metrics (PSA, CS, PET, BRS, PRS) are computed from these runs; see the paper for their definitions.

Layout

README.md
index.parquet                      one row per probe round
scores.parquet                     one row per probe of a round
SHA256SUMS                         sha256 of every archive
traces/<user>/<pa_model>__<condition>.tar.gz   one archive per setting

Each archive holds one setting (user x PA model x condition): its rounds and, for a memory condition, its session phase. Member paths are relative to traces/, so extracting every archive into traces/ gives:

traces/<user>/<pa_model>/<condition>/
  session/                         memory conditions only
    run_plan.yaml                  frozen plan of the run that holds the session phase
    steps/<acc_NNN|evolv_NN>/      one folder per session
  round_<1|2|3>/
    run_plan.yaml                  frozen plan of the round
    resolved_gt.yaml               the ground truth the round was scored against
    steps/<pre_NN|fin_NN>/         one folder per probe
    judge_input/<probe>/           ground_truth.json, ix_selection.json, meta.json, scene.json
    judge_output/<judge_id>/<probe>.json   the judge's verdict on one probe
    metrics/<judge_id>.json        the scorer's round metrics, when it wrote them
    prs/<fin_NN>.json              the user simulator's self-report on one final probe

To extract one setting:

sha256sum --check --ignore-missing SHA256SUMS
tar -xzf traces/user_a/openai_gpt-5.6-sol__mem0.tar.gz -C traces
ls traces/user_a/openai_gpt-5.6-sol/mem0      # session  round_1  round_2  round_3

<pa_model> is the OpenRouter id with / replaced by _. Inside a session or round folder every file keeps the path it has in a run directory of the evaluation code, so the code's readers can open it.

A step folder holds:

File	Content
transcript.jsonl	Event stream of the session or probe: start, simulator turns, PA turns with their tool calls, live-judge calls, end
transcript.md	Readable rendering of the same session or probe
meta.yaml	Step metadata: kind, context, target attribute and setting
pa_toolcalls.json	The PA's tool calls, grouped by beat
state/tool_log.jsonl, state/state_diff.jsonl	The tool calls and their effect on the user's tool world

Not every step has pa_toolcalls.json or state/. In transcript.jsonl, the endpoint of a judge_call event is the live-judge tier that answered: primary (OpenRouter) or fallback (a self-hosted endpoint).

prs/<probe>.json holds, for each fin probe, the user simulator's self-report after the probe, on the dimension the probe tests (PRS): score (1 to 5) and reason, with source the record it was taken from. status is ok for a self-report (7996), rule_assigned for a floored probe, scored 1 without asking (67), and rebuild_mismatch for a probe whose conversation the self-report could not rebuild exactly, so it was not asked (1; no score, and its round's PRS is the mean of the other fin probes).

index.parquet

Column	Type	Content
persona	string	user_a to user_d
pa_model	string	OpenRouter id of the PA model
memory_condition	string	One of the six conditions above
repeat	int8	Probe round, 1 to 3
run_id	string	Run id of the round, as in its run_plan.yaml
source_run_id	string	Run whose checkpoints the round's probes start from
judge_model	string	Offline judge model
round_status	string	standard, accepted_partial, diagnostic or incomplete (below)
round_status_reason	string	Reason recorded with a non-standard status, else null
floored_probes	list of string	Probes scored at the floor (below)
n_floored_probes	int8	Length of floored_probes
retried_probes	list of string	Probes whose shipped execution is a retry (below)
prs	float64	The round's PRS: the mean score of its prs/ files with status ok or rule_assigned
archive	string	The setting archive that holds the round
round_dir	string	The round's folder inside the archive
metrics_file	string	The round's metrics file inside the archive, or null when the scorer wrote none
session_dir	string	The session folder of a memory condition inside the archive, else null
session_skipped_steps	list of string	Sessions of that session phase that were skipped (below)

scores.parquet

One row per probe of each round, with the per-probe values the paper's round metrics are computed from. A round's PSA is the share of its 28 fin rows with setting_check_passed, and its BRS the mean brs of those rows. CS and PET combine the setting_check_passed of its rows as defined in the paper. Its PRS is the mean prs of its fin rows with prs_status ok or rule_assigned. The export checks that these give every round value of the paper's results.

Column	Type	Content
persona	string	As in index.parquet
pa_model	string	As in index.parquet
memory_condition	string	As in index.parquet
repeat	int8	As in index.parquet
run_id	string	As in index.parquet
probe_id	string	pre_01 to pre_05 or fin_01 to fin_28
context	string	work or personal
attribute	string	Preference attribute the probe tests
true_setting	string	True setting of that attribute in that context at the probe, as in the round's run_plan.yaml
setting_check_passed	bool	Whether the PA selected true_setting, from the probe's judge output; false for a floored probe
brs	int8	Behaviour score (BRS, 1 to 5) from the probe's judge output; 1 for a floored probe
prs	int8	fin probes: score of prs/<probe>.json (PRS, 1 to 5), null for rebuild_mismatch; pre probes: null
prs_status	string	fin probes: status of prs/<probe>.json; pre probes: null
floored	bool	The probe is in the round's floored_probes
retried	bool	The probe is in the round's retried_probes

Round status, floored, retried and skipped

All 288 rounds enter the paper's results, each with its 28 fin probes.

• round_status is standard for 269 rounds. The others were accepted with a caveat when the paper's results were compiled, and round_status_reason gives the recorded reason: accepted_partial (12), diagnostic (6), incomplete (1).
• floored_probes: 70 probes in 58 rounds have no usable execution after their attempts. They have no judge output and are scored at the floor: the setting check counts as missed and the behaviour score is 1. Their step folder holds what the last attempt left, if anything.
• retried_probes: for 10 probes the round's own attempt failed and the probe was run and judged again on its own; the step, judge input and judge output shipped for them are those of the retry.
• session_skipped_steps: 121 sessions across 42 session phases were skipped after failed attempts. Their step folder holds what the attempt left, if anything.

What is left out

Absolute paths of the machines that ran the benchmark are replaced by <run_root> (the root of the evaluation-code checkout, holding workspace/) or, for a path cut short in a preview, <host_path>. /home/... paths in the conversations are document paths inside the users' tool worlds. Provider reasoning and the judge's usage records are removed. Not included: the PA's full model context, memory-system inputs and outputs, checkpoints, tool-world snapshots, and run bookkeeping.

Ground truth

Ground truth is included by design, as in the user dataset: resolved_gt.yaml, judge_input/<probe>/ground_truth.json and the probe steps of run_plan.yaml carry each probe's true setting.

Anonymization

The four users are anonymized personas derived from episode transcripts of two U.S. television sitcoms. Their preference settings were extracted from the transcripts and checked by the authors. Names, employers, places and relationships were replaced through human-approved substitution tables, and quoted dialogue and catchphrases were removed. Their worlds, timelines, tool data and sessions were generated with LLMs and checked before acceptance. The dataset contains no source transcripts and no quoted dialogue.

In user_c's probe fin_17, a source first name was replaced by "User C", as in the personas dataset.

Citation

@misc{li2026mempabench,
  title         = {Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences},
  author        = {Li, Jialin and Chen, Zhenhao and Luo, Hanjun and Salam, Hanan},
  year          = {2026},
  eprint        = {2610.30485},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.30485}
}

License

The data is released under CC-BY-4.0. The evaluation code is released separately under the MIT license.