mem-agent: Persistent, Human-Readable Memory Agent Trained with Online RL

Source: https://huggingface.co/papers/Author: HuggingFace research Date: 2025-09-13

Summary

Paper presenting mem-agent: a 4B LLM trained to use an Obsidian-style markdown memory system via online RL. The agent learns to write, read, and update its own memory files to solve tasks that require persistent state.

Key Claims

  • Architecture: 4B language model + file system access (read/write markdown files). Memory is human-readable and inspectable.
  • Training method: online RL where the reward is task success — the agent learns which things to remember without being told.
  • Key finding: the agent learns “memory hygiene” (what to write, when to update, when to delete) from task outcomes alone.
  • Human-readable memory is both the feature and the constraint: the agent can’t use memory formats a human couldn’t understand.
  • Performance: outperforms both no-memory baselines and fixed memory (RAG) on long-horizon tasks.
  • Obsidian format specifically: markdown with wikilinks allows the agent to create a graph of memories, not just a list.

Connection to Other Sources

Directly relevant to Letta Context Repositories (different approach: file-based vs. git-based). Both are trying to solve agent memory with human-readable formats.

Concepts