NevaMind-AI/memU

What is memU?

memU (pronounced mem‑you) is a lightweight, agent‑driven personal memory system. It lets you keep a shared, searchable wiki of your own LLM‑generated knowledge that persists across:

different chat sessions,
multiple AI agents (ChatGPT, Claude Code, Cursor, OpenClaw, Hermes, WorkBuddy, etc.),
and even different devices.

The core of memU is only ~500 lines of Python, making the memory logic easy to read, audit, and extend.

Why use it?

Never lose useful context again – memU automatically captures the useful parts of each agent interaction and stores them as reusable Markdown “skills”.
Cross‑agent reuse – a skill learned while working with one agent (e.g., Claude Code) can be retrieved by another (e.g., Cursor) without any extra work.
Self‑hosting option – you can keep everything private on your own machine or use the hosted memU Cloud service.
Zero‑LLM overhead in the memory layer – memU only stores, embeds, and retrieves the Markdown you (or the agent) produce; it never makes its own LLM calls.

How does it work?

Host adapters – a small binary (e.g., memu-codex, memu-claude-code) knows where a given agent writes its session logs and which instruction file the agent reads.
Memorize – a scheduled background task reads new session logs, slices them into self‑contained jobs, and hands them to the agent.
Skill extraction – the agent decides whether to create a new skill, patch an existing one, or do nothing. The result is a clean Markdown file describing a reusable workflow.
Commit – memU writes the skill file back to the shared store, embeds the skill name/description, and indexes it.
Retrieve – before the agent answers a new query, memU injects a retrieve call that returns the most relevant skill(s) based on vector similarity.

All embedding and vector‑search work is handled by a configurable backend (in‑memory, SQLite, or Postgres with pgvector). The only LLM involvement is the agent that writes the skill markdown.

Quick start / Installation

# Install the CLI (includes the core library and all host adapters)
pip install memu-cli          # Python package
# or via npm/uv shortcuts
npx memu-cli --help
uvx --from memu-cli memu

Get an API key from the memU Cloud site https://memu.so (or set up your own embedding provider for self‑hosted mode).
Send a one‑line message to your agent (ChatGPT, Claude Code, etc.):
```
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is memu_•••••••••
```
The skill file guides the agent through detecting the host, installing the appropriate binary, configuring ~/.memu/config.env, and scheduling the bridging task.
Verify with <binary> doctor – it reports the resolved configuration and a successful retrieval test.

Using memU

Automatic retrieval – once installed, the host adapter injects a retrieve step before every agent response, so you get relevant skills without extra commands.

Manual retrieval – you can ask for a skill directly:

memu-codex retrieve "What should I remember about this project?"
# replace the binary with the one matching your agent

Developer integration – use memu memorize in your own applications to turn arbitrary conversation logs into memU‑compatible skill markdown. See docs/developer.md for the required JSONL format.

Ecosystem & Compatibility

OS	Agent	Memorize	Retrieve
macOS	ChatGPT (Work mode), Claude Code, Cursor, OpenClaw, Hermes, WorkBuddy	✅	✅
Windows	Same as macOS (with a few model‑specific notes)	✅	✅
Linux	Codex (VS Code), Claude Code, OpenClaw	✅/❌	✅

Supported adapters include:

memu-codex
memu-claude-code
memu-cursor
memu-openclaw
memu-hermes
memu-workbuddy
memu-cola
memu-pi
generic memu-agent for any other JSONL‑based agent.

Adding a new host only requires implementing a TranscriptSource (where the logs live) and a small CLI wrapper – the core pipeline is shared.

Configuration

All settings are read from environment variables, falling back to ~/.memu/config.env:

MEMU_DB – path/DSN of the vector store (SQLite default).
MEMU_EMBED_PROVIDER – openai, jina, voyage, doubao, openrouter.
MEMU_API_KEY – provider‑specific key.
MEMU_EMBED_MODEL – model name (defaults per provider).
MEMU_MEMORY_MODE – local (default) or cloud.

Run <binary> doctor to see the resolved configuration and test a retrieval.

License

Apache‑2.0 (permissive, commercial‑friendly).

TL;DR

memU gives you a personal, agent‑agnostic wiki that automatically learns from your LLM sessions and makes that knowledge instantly available to any supported AI coding assistant, all with a tiny, inspectable codebase.