How OpenClaw remembers on its own
OpenClaw is an open source, MIT-licensed personal AI assistant from the OpenClaw Foundation. It runs as a self-hosted gateway that connects chat apps such as WhatsApp and Telegram to AI agents. Its memory has no hidden state: everything it remembers is a Markdown file in the workspace (by default ~/.openclaw/workspace).
| File | What it holds | When the agent sees it |
|---|---|---|
MEMORY.md |
Long-term, curated facts and decisions | Loaded at the start of a session |
USER.md (optional) |
Stable preferences and profile of the user | Loaded with its own small budget |
memory/YYYY-MM-DD.md |
Detailed daily notes | Indexed for search, not injected every turn |
DREAMS.md (optional) |
Consolidation summaries for a human to review | For review |
The agent reaches this memory through tools from the active memory plugin, by default memory-core: memory_search finds notes by meaning, memory_get reads a file or line range. With an embedding provider configured, search is hybrid, vectors plus keywords. Two background habits keep it useful. Before compaction summarizes a long conversation, a silent turn asks the agent to write important context to durable memory. And “dreaming” promotes qualifying items from short-term recall into MEMORY.md. Both are on by default.
What the Mem0 plugin changes
The @mem0/openclaw-mem0 plugin (version 1.1.0 in Mem0’s documentation) replaces the memory slot. Instead of the agent writing notes to files, Mem0 handles three jobs:
- Triage: after each turn it extracts durable facts from recent messages and earlier work summaries, not from every message.
- Auto-recall: before each turn it retrieves the relevant memories and injects them into the context.
- Tools: the agent gets
memory_search,memory_add,memory_get,memory_list,memory_updateandmemory_delete, plus two event tools in platform mode.
Setting it up:
- Install with
openclaw plugins install @mem0/openclaw-mem0. - Activate it in
openclaw.jsonwith"plugins": { "slots": { "memory": "openclaw-mem0" } }. - Choose a mode. Platform mode (the default) sends conversations to
api.mem0.aiand needs a Mem0 API key. Open source mode needs no Mem0 key and runs on your own models and vector store. - Run
openclaw mem0 init; auto-recall, auto-capture and skills mode are then on, and each can be switched off.
Memories are scoped by userId, which defaults to your operating system username. Session memories are scoped by run, long-term memories persist across sessions, and recall searches both, long-term first.
Built-in or Mem0: the honest difference
| Built-in memory | Mem0 plugin | |
|---|---|---|
| Where memory lives | Markdown files you can open and edit | Mem0 cloud, or your own vector store |
| Who decides what is stored | The agent writes notes; dreaming promotes them | Mem0 extracts facts after each turn |
| Recall | The agent searches when it chooses to | Relevant memories injected before every turn |
| Readable by a human | Yes, plain files | Through the memory tools, or the Mem0 dashboard in platform mode |
The misunderstanding: OpenClaw needs Mem0 to remember
Many guides present Mem0 as the way to give OpenClaw memory, as if it had none. It already remembers, in files you can read, fix and put under version control. What Mem0 adds is automation: you no longer depend on the agent deciding to write something down or to search for it. The trade is visibility. A file you can open is easy to correct; extracted facts in a store need tools to inspect. When something wrong gets remembered, that difference matters (see how to manage AI agent memory).
What this means for a team
OpenClaw is built as a personal assistant, and its memory follows that: one workspace, one MEMORY.md, and in Mem0 one userId per namespace. Mem0’s documentation does not describe a general way to share memories across separate agents; subagents can recall their parent’s memories, and that is it. A colleague’s OpenClaw will not know what yours learned, and neither will the Claude or ChatGPT seat next to it. If a team wants one memory, the sharing has to live outside any single agent. For the wider landscape, see Mem0 vs Graphiti vs Cognee vs Honcho and How can a team share one AI memory?
Where Bombos fits
OpenClaw and the Mem0 plugin give one agent a memory. Bombos is a plugin that puts one shared memory layer underneath all the AI a team uses: Claude, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents. It learns how your team works, draws on what already sits in mail, documents and repos, and shows with every answer what it rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.
Related questions
Why does Mem0 not work right after installing the plugin?
OpenClaw treats memory plugins as an exclusive slot, so installing is not enough. You also set plugins.slots.memory to "openclaw-mem0" in openclaw.json, using the plugin ID rather than the npm package name.
Can I use the Mem0 plugin without sending data to Mem0?
Yes. Set mode to "open-source". No Mem0 key is needed then; by default it uses OpenAI models, so you need an OpenAI key, or you configure Ollama for a fully local setup with a vector store such as Qdrant or PGVector.
Which OpenClaw version does the Mem0 plugin need?
Mem0's documentation says OpenClaw 2026.4.25 or later is fully supported. Check yours with openclaw --version.
Are there other memory plugins for OpenClaw?
Yes. OpenClaw's own docs list the builtin SQLite-based memory as the default, plus Honcho and LanceDB as plugins. The bundled memory-wiki plugin adds a knowledge layer next to the active memory plugin instead of replacing it.
