The four simulated users of MemPABench, user_a to user_d, with every file the evaluation code reads to run and score them. The benchmark always runs these four users; the data does not change between releases of the code.

Use it through the evaluation code, https://github.com/mempabench/mempabench: uv run mempabench data pull downloads the revision pinned in the code and puts each file in place.

The benchmark is described in the paper at https://arxiv.org/abs/2610.30485.

Layout

One folder per user. Inside a folder, each file keeps the path it has relative to the root of the evaluation code:

Path inside `<user>/`	Content
`data/personas/<user>_identity.yaml`	Identity of the user
`data/personas/<user>_preferences.yaml`	Interaction-preference matrix: 14 attributes in the `work` and `personal` contexts
`data/state_fixtures/<user>/**`	Initial state of the user's tool world
`simulator/profiles/<user>/TOOLS.md`	Simulator profile seed
`Authoring/outline/drafts/<user>/outline.yaml`	Outline of the user's session timeline
`Authoring/sessions/normal/drafts/<user>/acc_NNN.yaml`	94 everyday session scripts
`Authoring/sessions/forced/drafts/<user>/evolv_NN.yaml`	5 preference-change events
`Authoring/sessions/forced/transcripts/<user>/evolv_NN/transcript.yaml`	The fixed transcripts of those 5 events
`Authoring/probe/outline/<user>/outline.yaml`	Probe schedule
`Authoring/probe/outline/<user>/resolved_gt.yaml`	Probe ground truth
`Authoring/probe/drafts/<user>/<probe_id>/<probe_id>.yaml`	33 probe scripts (5 `pre_NN`, 28 `fin_NN`)
`Authoring/probe/drafts/<user>/<probe_id>/fixture/**`	Changes a probe lays over the tool world; the empty `.gitkeep` files are part of it

Why the paths start with `Authoring/`

The directory names (Authoring/, drafts/) come from the pipeline that wrote the data; the files are the final versions used in the paper's runs, except that three user_c files had a source first name replaced by "User C". The paths stay as they were because the evaluation code reads each file at <code root>/<path>, records these relative paths and each file's sha256 in every frozen run plan, and each forced-event transcript names its source event by this path. Keeping them means data pull copies a user folder's contents into the code root unchanged, and runs record the same paths and digests as the paper's runs, except the digests of those three files.

Ground truth

Ground truth is included by design. Runs are scored on your own machine, so the preference matrices and each user's resolved_gt.yaml ship with the data.

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.

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}
}