How to give your OpenClaw agent memory that survives restarts

One of the first things people hit with OpenClaw is the "goldfish problem" — the agent feels sharp inside a session, then forgets everything the moment the daemon restarts. The fix is built in: MEMORY.md.

Here's the idea. MEMORY.md is a plain markdown file your agent reads on every session and writes back to as it learns. It's not a database, it's not a vector store — it's a file you can open, edit, and version-control like any other.

Step 1. Create the file.

Drop a MEMORY.md inside the directory your agent runs from. Empty is fine to start. The agent will pick it up on the next session.

Step 2. Seed it with what you actually want remembered.

Things that earn their place in MEMORY.md: who you are, what tools you use, recurring projects, names of people the agent will see again, conventions you don't want to repeat ("I prefer bullet summaries", "always reply in British English"). Things that don't: one-off task context, secrets, anything you'd rather it forget.

Step 3. Let the agent edit it back.

The whole point is durability across restarts, so let the agent append to it during sessions. If you want tighter control, review the diff at the end of the day and trim. Treat MEMORY.md like a journal you both write in.

A common mistake is making it too long. If MEMORY.md grows past a few hundred lines, the agent spends tokens re-reading background instead of doing work. Prune aggressively — keep it lean.

Hope that helps.

Cheers

Jason West 🙌

1 comment