English | 简体中文
Local-first hybrid memory system for OpenClaw.
- Structured layers: facts / preferences / tasks / episodes
- Pending turn queue:
turn_candidates.json - Rule-based post-turn classifier
- Queue-only autonomous lifecycle wiring (default off)
- Batch governance that can absorb safe drafts automatically
- Dedupe / merge / supersede / noop
- Fresh-workspace smoke path passes
flowchart TD
A[OpenClaw chat / sessions / cron] --> B[Turn capture via agent_end\nqueue-only when enabled]
B --> C[Post-turn classifier]
C --> D[Pending queue\nturn_candidates.json]
D --> E[Batch governance]
E --> F[Candidate drafts / preview / merge / supersede / noop]
F --> G[Structured memory\nfacts / preferences / tasks / episodes]
G --> H[Graph + reports]
G --> I[Exact search index]
J[memory-lancedb-pro\nsemantic recall layer] --> K[Semantic recall / hybrid memory]
A --> K
I --> L[Deterministic lookup]
H --> M[Admin console / audits]
K --> M
L --> M
| Capability | Purpose | Default | Dependency | Status |
|---|---|---|---|---|
| Structured memory layers | Store stable facts, preferences, tasks, episodes | On | None | Ready |
| Pending turn queue | Buffer new memory candidates safely | On | None | Ready |
| Queue-only lifecycle capture | Capture turns without direct structured writes | Off | None | Ready |
| Batch governance | Absorb safe drafts and refresh graph/reports | On | None | Ready |
| Exact search | Deterministic lookup over indexed memory | Optional build | None | Ready |
| Semantic recall | Higher-quality recall across phrasing variants | External | memory-lancedb-pro |
Recommended companion |
| Dedupe / noop / merge | Prevent noisy re-writes and duplicate intake | On | None | Ready |
| Fresh workspace smoke | Validate bootstrap → queue → governance flow | Manual/CI | None | Ready |
autoTurnCapture = falseautoTurnQueueOnly = trueturnCaptureMinConfidence = 0.88batchGovernanceEnabled = truebatchGovernanceEvery = 6h
Default principle:
queue first, govern second, absorb third.
The system does not directly write every captured turn into structured memory by default.
For full semantic recall, install and enable memory-lancedb-pro alongside this plugin.
Recommended tested stack for v0.1.1:
openclaw >= 2026.3.12claw-memory-system = 0.1.1memory-lancedb-pro >= 1.1.0-beta.8
Recommended install path:
openclaw plugins install memory-lancedb-pro
openclaw plugins enable memory-lancedb-proIf the current environment does not expose memory-lancedb-pro via the default plugin source, install it from its repository source instead, then enable it:
openclaw plugins install https://github.com/CortexReach/memory-lancedb-pro
openclaw plugins enable memory-lancedb-proWithout memory-lancedb-pro, this project still provides structured memory, queueing, governance, and exact search, but semantic recall quality will be significantly reduced.
- Install and enable
memory-lancedb-pro - Install and enable this plugin
- If your OpenClaw environment uses explicit plugin allowlists, add both plugins to
plugins.allow - Run bootstrap
- Enable or keep the batch governance cron
openclaw plugins install memory-lancedb-pro
openclaw plugins enable memory-lancedb-pro
openclaw plugins install <claw-memory-system-github-url>
openclaw plugins enable claw-memory-systemThen bootstrap the runtime:
Call claw_memory_bootstrap
- Install and enable
memory-lancedb-pro - Install and enable this plugin
- If your OpenClaw environment uses explicit plugin allowlists, add both plugins to
plugins.allow - Run bootstrap
- Build the exact index (optional)
- Enable or keep the batch governance cron
- Only enable lifecycle auto capture explicitly when you want queue-only capture
Recommended allowlist example:
{
"plugins": {
"allow": [
"memory-lancedb-pro",
"claw-memory-system"
]
}
}docs/autonomous-memory-runtime.mddocs/autonomous-memory-runtime.zh-CN.mddocs/quickstart-openclaw-chat-install.zh-CN.mddocs/full-enable-guide.zh-CN.mddocs/release-notes-v0.1.zh-CN.mddocs/final-release-matrix.zh-CN.md
Historical implementation plans have been moved under docs/archive/plans/ to keep the release surface clean.