Add Memory to OpenClaw¶

OpenClaw gets the MCP tools. It does not get automatic injection or automatic capture. That is the current state, stated plainly rather than buried: on OpenClaw this is instructed memory, not subconscious memory.

Pre-release

popoto[mcp] is not on PyPI yet. Until a release ships, install from a checkout: pip install -e '.[mcp]'.

Install¶

pip install 'popoto[mcp]'
popoto-memory doctor

Add to ~/.openclaw/openclaw.json under mcp.servers:

{
  "mcp": {
    "servers": {
      "popoto-memory": {
        "command": "popoto-memory",
        "args": ["mcp"]
      }
    }
  }
}

Restart OpenClaw. You get four tools:

Tool	Purpose
`memory_search`	Find something stored earlier
`memory_save`	Store a fact
`memory_feedback`	Mark a memory `contradicted` or `acted`
`memory_status`	Connection, scope, retrieval mode, record count

What you are missing, and why¶

On Claude Code, Codex, and Hermes, memory runs on every turn because the harness fires a hook that this package handles. The model is never asked. That is the property that makes memory reliable: a memory system is only worth having if things were actually recorded to recall later, and a model that must choose to record will not choose consistently.

On OpenClaw the model has to call memory_save and memory_search. It will sometimes. It will often not.

OpenClaw's per-turn hooks -- before_prompt_build, which returns appendContext, and llm_output -- live in TypeScript plugins rather than in a command string. The adapter here already speaks that contract: tests/fixtures/harness_payloads/openclaw_*.json round-trip through it and produce {"appendContext": "..."}. A thin plugin that shells out to popoto-memory hook is all that is missing.

Whether an OpenClaw plugin may spawn a subprocess is unverified. Plugins load as in-process Node modules, so child_process should be reachable, and no documentation says otherwise. "Should be" is not verification, and OpenClaw is not installed on any machine this integration is developed on, so the plugin is not shipped.

The alternative -- reimplementing the memory path in TypeScript -- would be a second implementation of the core, and is not going to happen.

Resolving this takes one person with OpenClaw installed running a plugin that calls child_process.execFile. If that works, the plugin is a dozen lines: pass the event through to popoto-memory hook, return its JSON.

Getting more out of the tools in the meantime¶

Because recall is model-elected here, telling the agent when to search actually helps, which is not true on the other harnesses. Something like this in your instructions:

Before answering a question about how this project works, call
memory_search. After learning something durable about the project, call
memory_save.

That is a workaround for a missing hook surface, not the intended design.

Verify¶

popoto-memory doctor

records climbing means memory_save is being called. If records stays at 0 across a long session, the model is not electing to use the tools -- which is the whole problem described above, not a bug in the setup.

Configuration¶

Identical to every other harness, all environment-driven. See Harness Integration.