Recall for OpenClaw

Long-term memory for OpenClaw agents, kept in Recall by Polign. The plugin fills OpenClaw's memory slot: before each turn it puts what Recall remembers in front of the prompt, after each turn it sends the conversation to Recall, and the agent gets three tools to search, store and forget on its own.

What Recall adds over a search index:

• A correction replaces. "I moved to Bellevue" replaces "I live in Seattle". The agent sees the current value and what it replaced, so it can mention the change when that matters, and it never sees both as equals.
• Nothing is deleted. Every statement stays in the history. memory_search takes as_of, so the agent can ask what was believed last month.
• Your server, your bucket. Memories go over HTTP to recall serve, which keeps them in polign_db on local disk or in your own object store bucket. The Recall key chooses the namespace, so one server can hold memory for many agents without any of them reaching another's records.

Before you start

• OpenClaw 2026.10.2 or later on Node 24.16 or later.
• recall 0.16.0 or later: brew install polign/tap/recall or pip install polign-recall. Either way you also get polign-server.
• A model for working facts out of text, as provider:model: a hosted one such as anthropic:claude-haiku-4-5-20251001 with its key in the environment, or a local one such as ollama:qwen3:8b. Without one, Recall keeps captured text as written: searchable, but nothing replaces anything, and memory_forget is refused. A small local model misses some phrasings; in testing, qwen3:8b caught "I live in Bellevue now" but not "I moved to Bellevue last week".

Run the memory server

Start polign_db and recall serve in text mode, then mint a key for OpenClaw. The server holds the database; OpenClaw holds only its key.

polign-server -store fs:$HOME/.local/share/polign/openclaw
recall -url http://127.0.0.1:23000 serve -text -listen 127.0.0.1:8080 \
  -extract-model anthropic:claude-haiku-4-5-20251001
recall -url http://127.0.0.1:23000 key create -namespace openclaw -expires 8760h

The last command prints a key that starts with rcl_, once. That is the only credential the plugin needs; the namespace it names is where the memories live. Both servers listen on loopback by default. -store takes an S3, GCS or Azure URL as well as a local path, and a database that requires keys takes its key in POLIGN_API_KEY; see Run polign_db.

Install the plugin

openclaw plugins install @polign/openclaw-recall

OpenClaw asks you to confirm, because the package is not yet traced to a signed build. In a script, add --force --accept-capabilities. The local example does all of the steps below for you.

Then in openclaw.json:

{
  "plugins": {
    "slots": { "memory": "recall" },
    "entries": {
      "recall": {
        "enabled": true,
        "hooks": { "allowConversationAccess": true },
        "config": {
          "url": "http://127.0.0.1:8080",
          "apiKey": "${RECALL_API_KEY}"
        }
      }
    }
  }
}

slots.memory makes Recall the memory engine instead of the built-in one. hooks.allowConversationAccess lets the plugin read the prompt and the turn's messages, which auto-recall and auto-capture need; OpenClaw asks this of every plugin it did not ship. Put the key in RECALL_API_KEY or write it in place of ${RECALL_API_KEY}.

Then turn off the built-in memory plugin. OpenClaw 2026.10 keeps it loaded even when the slot points elsewhere, and it registers memory_search first, so without this step the agent's searches go to the workspace Markdown files instead of Recall:

openclaw plugins disable memory-core

Restart the gateway, then:

openclaw recall status
openclaw recall remember "I moved to Bellevue and I prefer concise answers"
openclaw recall search "where does the user live?"

To try the plugin from a checkout instead of npm, build it and point OpenClaw at the folder:

cd /path/to/polign/plugins/openclaw && npm install && npm run build
openclaw plugins install --link --force --accept-capabilities "$PWD"
openclaw plugins inspect recall --runtime

What the agent gets

And two hooks, each on by default:

• Auto-recall (before_prompt_build): asks Recall about the user's message and puts up to maxMemories answers in front of the prompt, inside a <relevant-memories> block. Nothing is added when Recall has nothing to say or does not answer.
• Auto-capture (agent_end): after a successful turn, sends the latest captureMessages user and assistant messages to Recall as one passage. Injected memory blocks are stripped first, so memory never feeds itself. The capture runs in the background, one at a time, because extraction can take longer than the 30 seconds OpenClaw waits for a hook. It needs a process that stays up, such as the gateway: a one-shot openclaw agent --local exits before the capture finishes.

Settings

All under plugins.entries.recall.config. The environment is the fallback for the first two.

What works differently from the built-in memory

• Memories are statements Recall worked out, not Markdown files in the workspace. MEMORY.md and the daily notes are not read or written.
• There is no list-everything tool. openclaw recall search with a broad question, or recall explain against the server, show what is held.
• Recall needs its server running. When it is down, the hooks log a warning and add nothing, and the tools report the error to the agent.

Develop

npm install
npm test          # node --test, no network
npm run check     # tsc against the OpenClaw plugin SDK
npm run build     # dist/, the compiled entry OpenClaw loads from npm

From a checkout OpenClaw loads the TypeScript source directly; from npm it loads dist/index.js, which npm publish builds first. openclaw plugins validate applies to plugins written with defineToolPlugin, not to this one; use openclaw plugins inspect recall --runtime to see what it registered.

Source and license

Apache License 2.0, in github.com/Polign/polign/plugins/openclaw. Recall itself is at github.com/Polign/recall, and the HTTP contract the plugin is written against is its docs/memory-api-v1.md.