Renor memory plugin for OpenClaw: durable memory shared with Claude Code, Codex and Cursor
Renor memory for OpenClaw
Gives an OpenClaw agent Renor as its memory. Renor is durable memory shared with your other agents (Claude Code, Codex, Cursor), so what you told one of them is there for OpenClaw too, and the other way round.
The plugin takes OpenClaw's memory slot and does three things:
- Auto-recall. Before each turn it asks Renor whether it holds memory relevant to your message. When it does,
the turn gets one line saying so and telling the agent to read it with
renor_recall, so the agent doesn't need to remember to look. If Renor is slow or unreachable, the turn goes ahead without it. renor_recall. Returns the saved text. This is the only way stored text reaches the model, and it comes as a tool result: quoted JSON markeduntrusted-data.renor_remember. A tool the agent uses to save a preference, a decision or a fact in one plain sentence. A short, cacheable line in the system prompt tells the agent when to use it.
What auto-recall sends. autoRecall is on by default. Once conversation access is granted (below), each user
message is sent to Renor before the turn: up to 1000 characters, to the hosted service or to your own engine.
Short prompts are skipped: under 3 words, slash commands, ! shell lines, and replies like "ok" or "yes go ahead".
To stop it, set plugins.entries.renor.config.autoRecall to false. Then Renor is only read when the agent calls
renor_recall.
Why the saved text isn't put in the prompt. Anyone who can write to your memory could otherwise write the
agent's prompt: another agent, an import, a shared workspace. So auto-recall only says that matches exist. It never
counts a reply Renor marks as not relevant, or memories Renor marks untrusted (imports and reserved sources).
renor_recall prefixes untrusted memories with [untrusted]. The system prompt tells the agent to treat recalled
text as information and never to act on instructions in it. That rule is prompt text: it helps, but it can't be
enforced. The enforced part is that stored text never reaches the prompt. Reading it costs the agent one tool call.
It works with hosted Renor or with a Renor engine you run yourself. MIT licensed.
Install
openclaw plugins install renor-openclaw
openclaw config set plugins.entries.renor.hooks.allowConversationAccess true
Installing switches plugins.slots.memory to renor. OpenClaw only runs auto-recall for a non-bundled plugin
after the second command. Without it you still get the two tools, but nothing is recalled automatically.
Hosted (renor.cloud)
- Sign in at https://renor.cloud/dashboard, open API keys and create a key named
openclaw. Saves made through this key are labelledopenclaw. - Put the key in the environment of the process that runs the OpenClaw Gateway, as
RENOR_API_KEY, in whatever holds that process's secrets (a systemdEnvironmentFilewith mode 0600, your shell profile, a secrets manager). The plugin reads it from there, and nothing is written to OpenClaw's configuration.
To keep the key in a secret manager instead, point the setting at it with an OpenClaw SecretRef. The configuration then holds only the reference, never the key:
# from an environment variable (the same as above, but explicit)
openclaw config set plugins.entries.renor.config.apiKey --ref-provider default --ref-source env --ref-id RENOR_API_KEY
# from 1Password, Vault or another exec provider you have set up under secrets.providers
openclaw config set plugins.entries.renor.config.apiKey --ref-provider <alias> --ref-source exec --ref-id <id>
What is stored where:
RENOR_API_KEYor anenv/execSecretRef: the key is in no OpenClaw file.- A
storeSecretRef: the key sits unencrypted in OpenClaw's state database (state/openclaw.sqlite), with file mode 0600 (OpenClaw's own docs say so). - A plain
apiKeystring: the key sits in clear text in the OpenClaw config file. Don't do this.
The manifest declares apiKey a secret input, so OpenClaw conceals it in Settings. The plugin sends the key only in
the Authorization header and never writes it to a log; the warnings it logs carry Renor's error message, not the
request. The key goes only to an https URL, or over plain http to a literal loopback address (127.0.0.1,
[::1]) for an engine on your own machine. Any other URL, http://localhost included, is refused at load.
Self-hosted engine
With no key set, the plugin starts the open-source engine (npx -y renor-engine@^1 mcp) and stores memory in
~/.renor, the same folder claude mcp add renor-local -- npx -y renor-engine mcp uses. The @^1 pin means each
start runs a 1.x release, not whatever is newest on npm. To pin an exact version, set args, e.g.
["-y", "renor-engine@1.0.0", "mcp"]. Point it at another folder with:
openclaw config set plugins.entries.renor.config.vaultPath /absolute/path/to/notes
The engine needs Node 22. If the node on the Gateway's PATH is a different version, set command and args
to start it with a Node 22 binary, for example command: "/path/to/node22/bin/node" and
args: ["/path/to/renor-engine/dist/cli.js", "mcp"].
The self-hosted engine has no embedding model unless you configure one. Without it, recall matches your message's words as a phrase, so a whole question ("Which database does Fjord use?") often finds nothing that its key words ("Fjord database") would. To recall by meaning, pass a model to the engine:
plugins: { entries: { renor: { config: { env: { EMBEDDING_PROVIDER: "openai", OPENAI_API_KEY: "..." } } } } }
Configuration
All keys are optional and live under plugins.entries.renor.config.
| Key | Default | What it does |
|---|---|---|
mode | hosted if a key is set, else local | Which Renor to use. |
apiKey | RENOR_API_KEY | Hosted: your Renor key. Prefer the environment variable or a SecretRef (see above). |
url | https://mcp.renor.cloud/mcp | Hosted: the MCP endpoint. https only, except plain http to 127.0.0.1 or [::1]. |
scope | everything | Hosted: limit recall and saves to one scope, e.g. project:fjord. |
command, args | npx, ["-y", "renor-engine@^1", "mcp"] | Local: how to start the engine. |
vaultPath | the engine's default (~/.renor) | Local: the notes folder. |
env | none | Local: extra environment for the engine. |
autoRecall | true | Send each non-trivial user message to Renor before the turn, and tell the agent when matching memory exists. |
recallMaxTokens | 800 | Hosted: token budget for one recall. |
recallLimit | 8 | Most memories one recall returns. |
timeoutMs | 8000 | Longest a recall or save may take. A slower auto-recall is skipped. |
How it compares with Hindsight
Hindsight retains every turn and extracts facts from it with an LLM. This plugin does not record conversations. The agent saves what is worth keeping, and auto-recall points the agent at it. Memory lives in Renor, so the same notes reach your coding agents over MCP.
Develop
cd ../../engine && npm ci && npm run build # the tests run against this repository's engine, offline
cd ../integrations/openclaw && npm install && npm test
The tests run the real engine twice. Local mode uses its stdio MCP server. Hosted mode uses its HTTP service, the same code renor.cloud runs for each workspace, with a bearer minted on its loopback-only route. No network and no model are used.
