Respire for OpenClaw

respire-openclaw installs the respire memory engine into OpenClaw as a native TypeScript plugin: title-tier recall injected before every turn, seven rsrs tools for the agent, judge-then-store receipts. See the plugin overview for the shared runtime behavior, first-use steps and tuning reference.

Install

# from a local checkout (dev):
git clone https://github.com/risense-ai/respire-openclaw.git
cd respire-openclaw && npm install && npm run build
openclaw plugins install --link ./respire-openclaw

# or from GitHub:
openclaw plugins install git:github.com/risense-ai/respire-openclaw

The engine (rsrs) is discovered automatically (npm i -g rsrs-cli@latest when missing — the plugin hands the command to the agent and recovers within the same session).

Enable conversation hooks

The per-turn recall hook is a prompt-modifying conversation hook, so OpenClaw requires explicit permission for non-bundled plugins. In openclaw.json:

{
  plugins: {
    entries: {
      respire: {
        enabled: true,
        hooks: { allowConversationAccess: true },
        config: {},
      },
    },
  },
}

Then verify with openclaw plugins inspect respire --runtime --json. Hook permission key names follow docs.openclaw.ai/plugins/hooks (checked 2026-11); if your version rejects the key, consult that page for the current schema.

What happens in a session

Moment	Behavior
First prompt of a session	Compact respire rules injected as system context, plus a one-shot engine probe: engine-missing / not-signed-in / empty-library / model-missing guidance (probe cached with status_ttl_ms)
Every turn (before_prompt_build)	Automatic rsrs recall --titles: title-tier injection — score + short id + title only, budgeted; slash-command tokens are stripped before querying; unfiltered by default
Store prefixes	「记：」/「存：」/「记住：」/remember:/store: → nothing is written by the plugin; a pending-placement handoff tells the model to judge (update > merge > attach > new) and store via the tools or the CLI
Agent tools	respire_recall / respire_show / respire_chain / respire_remember / respire_update / respire_attach / respire_status; important without a verified parent never writes
/nomem prefix	Recall skipped for the turn
Slash commands	/respire-recall <query>, /respire-status, /respire-sync

Enable and disable

Level	How	Scope
Single turn	/nomem prompt prefix	This turn's recall only
Recall switch	"recall_enabled": false in the plugin config	Tools/commands/rules stay, auto recall stops
Master switch	"enabled": false in the plugin config	Whole memory layer silent
Plugin level	openclaw plugins disable respire	OpenClaw-side full off

First use

1. Register at rsrs.rs, then run rsrs register on the gateway machine. Daily: rsrs login, rsrs status.

2. Import existing rules — ask the agent on any OpenClaw-connected surface:

   Import my existing rules and memories into respire: read, one by one — ~/AGENTS.md, the AGENTS.md at the roots of your usual projects, and ~/.agents/skills/*/SKILL.md; extract reusable facts, conventions, lessons and red lines (secrets, ID numbers, card numbers, private keys are always excluded — record only their location), deduplicate, then store them item by item with respire_remember (importance important with a verified parent, titles as fact sentences). Give me an inventory when done.

   An empty library gets the same nudge on day one, injected with the session rules.

3. Optional model: rsrs model install && rsrs model activate m3.

Tuning

Config: plugins.entries.respire.config in openclaw.json (validated against the manifest configSchema; reload the plugin after changes). Shared fields are documented in the overview. OpenClaw-specific keys: recall_enabled (auto recall off, tools and rules stay) and status_ttl_ms (probe cache lifetime).

Uninstall

openclaw plugins remove respire

Known limits

• No Stop-phase storage check — OpenClaw has no blocking end-of-turn hook, so the family's one-extra-round nudge does not exist here; the rules block carries the "store before finishing" duty instead. Store when the turn produced facts.
• No PreToolUse write guard — the write guards live inside the tools themselves. Raw rsrs remember calls from the agent's shell bypass the plugin guard by design; engine-side policy still applies.
• Real-gateway loading is not CI-verified: manifest and API shapes are statically aligned with docs.openclaw.ai (checked 2026-11); verify on your gateway with openclaw plugins inspect respire --runtime --json.
• The prompt-injection permission requires hooks.allowConversationAccess in your openclaw.json; without it the tools and commands still work but per-turn recall stays off.

Tests and design notes live in the repository README.