OpenClaw plugin

The OpenClaw adapter (@plur-ai/claw) is the deepest integration PLUR has anywhere. It plugs into OpenClaw’s ContextEngine — the exclusive plugin slot with seven lifecycle hooks — so PLUR can shape every stage of the conversation: bootstrap, ingest, assemble, compact, after-turn, subagent spawn, and subagent end.

The result: memory works without the agent ever calling a tool. No system prompt nudging required.

Install

openclaw plugins install @plur-ai/claw
openclaw gateway --force
openclaw config set plur.enabled true

That’s the whole setup. The plugin is auto-discovered, its hooks register, and memory is live from the next session.

What the seven hooks do

Hook	PLUR action
bootstrap	Open the local engram store, warm the index.
ingest	Scan incoming user content for correction patterns; stage candidate engrams.
assemble (required)	Inject relevant engrams into the assembled system prompt, before the LLM call.
compact (required)	When context fills, prioritise pinned/locked engrams; spill the rest.
afterTurn	Capture the turn as an episode; check for learnings the agent surfaced.
prepareSubagentSpawn	Hand the subagent a scoped engram bundle so it inherits memory.
onSubagentEnded	Merge anything the subagent learned back into the parent store.

The contract is documented in OpenClaw’s PR #22201 (Josh Lehman); @plur-ai/claw implements all seven. Most plugins implement only assemble and compact and miss the rest of the loop.

What you get vs MCP

	MCP adapter	OpenClaw plugin
Injection requires agent tool call	yes (or hooks if on Claude Code)	no — automatic
Captures corrections from raw user text	only via Claude Code hooks	yes — every adapter
Subagent memory inheritance	no	yes
Per-turn episode capture	manual	automatic
Compact-aware pinning	no	yes

If you’re choosing between MCP and OpenClaw for a project that already runs on OpenClaw, take OpenClaw. It’s strictly more capability.

Combined with mcporter

Inside OpenClaw, the plugin handles the invisible memory layer (hooks). For explicit tools the agent can call — plur_learn, plur_recall, etc. — pair @plur-ai/claw with mcporter bridging @plur-ai/mcp. You get hooks and tools:

openclaw plugins install @plur-ai/claw
openclaw mcp add plur npx -y @plur-ai/mcp

Configuration

The plugin reads ~/.openclaw/plur.yaml:

plur:
  enabled:           true
  store:             ~/.plur/             # path to engram store
  inject:
    max_engrams:     12                   # cap per assemble
    pinned_always:   true                 # always include pinned even if irrelevant
  capture:
    correction_patterns: [default]
    auto_promote_after: 3                 # candidate becomes active after 3 confirmations
  subagent:
    inherit_scope:   true
    bundle_size:     20

Defaults are sensible — most installs never edit this file.

Version compatibility

@plur-ai/claw	OpenClaw
0.7.6	2026.4.2+
0.7.5	2026.3.7–2026.4.1
0.7.0–0.7.4	2026.3.7

For older OpenClaw versions, check the release notes. The plugin uses OpenClaw’s pluginApi compatibility shim where needed.

Troubleshooting

• Plugin lists as installed but doesn’t fire — run openclaw gateway --force to rebuild the plugin registry.
• Subagent doesn’t inherit memory — confirm subagent.inherit_scope: true and that the parent session is using the same store path.
• High memory usage on assemble — lower inject.max_engrams; default 12 is generous.

Next: Hermes (Python) →