---
title: README
emoji: 🤖
colorFrom: yellow
colorTo: purple
sdk: static
pinned: false
---

<p align="center">
  <img src="https://raw.githubusercontent.com/mempabench/mempabench/45dea87/docs/assets/hero.gif" width="720" alt="A robot agent holds its notebook about the user, where &quot;Orders: just do it, don't ask&quot; is circled in red, yet it raises a sign asking &quot;Shall I proceed?&quot;. The user, shopping on a phone, face-palms and says &quot;Just do it!&quot;">
</p>

**Can personal agents keep up with your interaction habits across contexts and over time?** MemPABench tests whether an agent's memory lets it learn when each user wants it to ask, confirm, or just act.

| | |
|---|---|
| Paper | [arXiv:2610.30485](https://arxiv.org/abs/2610.30485) |
| Code | [github.com/mempabench/mempabench](https://github.com/mempabench/mempabench) |
| Website | [mempabench.com](https://mempabench.com) |
| The four simulated users | [mempabench-personas](https://huggingface.co/datasets/mempabench/mempabench-personas) |
| The paper's runs | [mempabench-traces](https://huggingface.co/datasets/mempabench/mempabench-traces) |