Agent memory as a filesystem that acts before you ask

What it does

memU gives AI agents a persistent, structured memory layer designed for 24/7 operation. It stores facts, preferences, and skills in a hierarchical file-system-like structure—categories act as folders, memory items as files, and cross-references as symlinks—so agents can navigate context without dumping entire conversation histories into the prompt. A parallel “MemU Bot” monitors interactions, extracts insights, and injects relevant context back into the main agent to keep responses cheap and targeted.

The interesting bit

The framework treats memory as a mountable, browsable filesystem rather than a flat vector database, which makes long-term knowledge portable and queryable by structure. It also attempts proactive behavior: the sidecar process predicts user intent and pre-fetches context before the next prompt arrives, though the README is truncated before detailing how that prediction actually works.

Key highlights

• Hierarchical memory organized like directories, files, and symlinks
• Parallel proactive bot that monitors, memorizes, and predicts intent alongside the main agent
• Claims to reduce long-running token costs by caching insights and avoiding redundant LLM calls
• Supports mounting external resources (conversations, documents, images) as queryable memory
• Requires Python 3.13+

Caveats

• The README is truncated, so deeper technical details (storage backend, exact retrieval mechanics, and the “self-evolving skills” mentioned in the repo description) remain unseen
• Aggressive Python 3.13+ requirement may limit deployment options
• The framework and its “enterprise-ready” sibling memUBot are presented interchangeably, which can blur the open-source boundary

Verdict

Worth a look if you’re building always-on agents and want memory that feels more like a filesystem than a database. Skip it if you need battle-tested, fully documented infrastructure today; the visible docs are heavy on diagrams and light on implementation specifics.