memU Gives Your Coding Agent a Memory That Outlives the Session

Every coding agent forgets you the moment the session ends. memU is a 14k-star, Apache-2.0 memory system that fixes this: it mines your agent's session history, distills the reusable parts into Markdown skills and memory, and injects the relevant pieces back before your agent answers — across sessions, agents, and devices.

Why This Skill Matters

Context windows reset, but your project knowledge should not. memU runs as a sidecar to desktop coding agents — the README lists Codex, Claude Code, Cursor, OpenClaw, Hermes, WorkBuddy, Cola, and pi among its tested integrations — and points them all at one shared memory backend. What your Codex session learned, your Claude Code session can retrieve.

Two design choices make it inspectable. The core memory logic is 500 lines, compact enough to read and adapt. And the judgment stays inside your agent: the MemoryService component makes no LLM or chat calls — it only stores, embeds, and retrieves the Markdown your agent prepared.

Installation

The README's quick start is one message to your agent. Get an API key from memu.so, then send:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is memu_•••••••••(get Api Key from memu.so).

To self-host instead — private, single-device, with your own storage and embedding provider — the README gives a parallel message pointing at the repo's SKILL.md:

Read https://raw.githubusercontent.com/NevaMind-AI/MemU/main/SKILL.md and follow it to install memU.

You can also install the CLI directly:

pip install memu-cli         # library + memu + memu-codex CLIs
npx memu-cli --help          # CLI via npm launcher (engine: PyPI package memu-cli)
uvx --from memu-cli memu     # CLI via uv, no install

Real Workflow: Let Claude Code Remember a Debugging Win

1. Send the quick-start message. The SKILL.md routes your agent through the install: identify the host, print that host's install guide, configure the memory backend, register a scheduled bridging task, and patch the instruction file — each step behind a verify gate.
2. For Claude Code, the memu-claude-code adapter mines ~/.claude/projects/<project>/<session>.jsonl and adds a standing instruction to ~/.claude/CLAUDE.md telling the agent to run retrieve before answering.
3. Verify the loop with the doctor command:

memu-claude-code doctor

4. Work normally. A scheduled task slices new session logs into self-contained jobs; your agent reads the related existing skills and then decides to do nothing, patch a skill, or create a new one — capturing useful branches, edge cases, and pitfalls as readable Markdown.
5. Weeks later, ask in any connected agent, or retrieve manually:

memu-claude-code retrieve "What should I remember about this project?"

The retrieval happens automatically before answering once the standing instruction is in place; the manual command is there when you want to inspect what memU knows.

Tips

• If the selected model declines the setup steps, the README's agent-support note says to retry with Opus or another model — Sonnet 5 can occasionally decline.
• Cloud mode is cross-device and free per the README; local mode needs an embedding provider (OpenAI, Jina, Voyage, Doubao, or OpenRouter) and a MEMU_DB store — SQLite by default, Postgres with pgvector for concurrent access.
• MEMU_MEMORY_MODE selects Local or Cloud; an unset mode stays Local for backward compatibility.
• Run the adapter's doctor subcommand after config changes — it prints the resolved mode and exercises a live retrieval.
• Uninstalling keeps your memory store and ~/.memu/config.env by default, so a reinstall resumes where you left off; memory is erased only when you explicitly ask.

When Not to Use This

The integration matrix marks Claude (Chat and Cowork) and ChatGPT's Chat mode as unsupported — memU targets coding agents with session logs it can mine. If your work is single-session and throwaway, a CLAUDE.md note is simpler. And if you would rather no third party see your history, use the self-hosted path instead of the cloud key.

See the leaderboard for more skills.