Memoir for OpenClaw: scoped memory that stays in its lane

OpenClaw is a personal-assistant gateway — one process that can front several agents and many concurrent conversations. That makes memory trickier than it is for a single chat window: whose facts are these, and which conversation should see them? The memoir plugin answers both. It fills OpenClaw's memory slot with a versioned, git-like store — and it gives that store a scope, so memory lands where it belongs and nowhere else.

It just remembers — across sessions

You don't manage memory; you talk. Tell OpenClaw your name and it captures the durable fact in the background after the turn. Start a fresh session later and it already knows — the fact is surfaced into the prompt automatically, no tool call required. Under the hood two tools do the work when the model needs them: memoir_recall to read stored facts (LLM-free, so it's cheap) and memoir_remember to write them, classified into a taxonomy and guarded against secrets.

Scoped memory: the part that matters for a gateway

Here's the feature worth slowing down for. Because OpenClaw can serve multiple agents and conversations, memoir lets you choose how wide a memory's reach is, with one config key: plugins.entries.memory-memoir.config.scope. Three modes:

• profile (the default) — one memory per agent (profile-<agentId>), shared across all of that agent's sessions. Your assistant remembers you from launch to launch, but one agent's memory never bleeds into another's.
• chat — one memory per conversation (chat-<sessionKey>). Every launch starts clean and nothing carries over — right for shared terminals, one-off threads, or anything you want walled off by default.
• off — a single shared default store. Every fact is globally visible. Simplest, no isolation — one unified profile.

The isolation is default-deny, and the mechanism is worth understanding. Captures and memoir_remember write only to the scoped namespace. Reads (memoir_recall) look at the scoped namespace plus the shared default namespace. So a fact learned inside a scoped conversation stays confined to it, while anything you deliberately place in default is visible everywhere. Under the default profile scope the plugin writes nothing to default — it stays empty — so there's no accidental global leakage to clean up later.

The practical payoff: a hypothetical you explore in one chat, or facts tied to one persona, can't contaminate every other conversation the gateway is handling. "This conversation's memory" becomes a real boundary instead of a hope.

Everything is still a commit

Scope decides where memory lives; memoir's git-like store decides what you can do with it. Every write is an inspectable commit, so the questions that bite long-lived assistants all have answers:

• What's stored, and where? memoir_status reports the branch, memory count, and current scope.
• Where did this come from? memoir log and memoir blame trace any fact to the session that wrote it.
• Take it back. memoir_forget deletes by taxonomy path; history stays recoverable.
• Branch and merge. /memoir branch, /memoir checkout, and /memoir sync diverge a timeline and fold it back when you choose.

All of it is reachable in chat through the /memoir command (status, recall, remember, forget), and openclaw memoir ui opens a visual browser over the whole store.

Getting started

It's a one-time setup on the gateway host:

• Install the memory tool: pip install memoir-ai
• Install the plugin: openclaw plugins install https://github.com/zhangfengcdt/openclaw-memoir
• Allow conversation access: openclaw config set plugins.entries.memory-memoir.hooks.allowConversationAccess true
• Restart the gateway: systemctl --user restart openclaw-gateway

Leave scope on profile for a personal assistant that remembers you between launches, or set it to chat when each conversation should stand alone. Verify any time with openclaw memoir status.

Set up memoir for OpenClaw

A gateway's memory shouldn't be all-or-nothing. Scope makes "remember this for me" and "keep this to this conversation" two different, deliberate choices — and the store stays inspectable either way.