Memory that cleans up after itself.
OpenClawBrain does not just help agents remember more. It runs passive graph maintenance so agents maintain what they remember: fewer duplicate memories, safer graph edges, scoped exceptions, stale-memory review, privacy-safe forgetting, and proof for every change.
Memory Authority decides what can influence this turn. Graph Maintenance curates how the memory graph evolves over time.
Why this matters
Agent memory gets messy. The same correction may be captured twice. An old workflow may stop matching the repo. A broad preference can conflict with a narrow task. A deleted memory can still leave unsafe edges behind.
Generic memory systems usually optimize for recall. OpenClawBrain optimizes for trustworthy continuity: remember useful things, but keep the graph healthy enough that old context does not silently overpower the present.
The two jobs
Memory Authority
Runs during a turn. It asks whether a retrieved memory is current, scoped, safe, non-superseded, private enough, and compatible with what the user just asked.
Graph Maintenance
Runs in the background after many turns. It proposes cleanup, lineage, edge repair, stale-memory review, scoped exceptions, and tombstone recapture blocking, while safe auto-apply stays limited to deterministic low-risk repairs.
The safety line is strict: graph connectivity is not authority. A maintenance proposal can provide features and proof, but Memory Authority still recomputes turn-level use.
What gets maintained
Duplicate memories
Exact duplicates can be consolidated into a canonical node while retaining lineage and proof.
Bad edges
Edges to missing, deleted, or tombstoned nodes can be retired instead of polluting retrieval.
Stale authority
Old high-authority memories are flagged for review before they keep steering future work.
Scoped exceptions
Broad defaults and narrow instructions stay separate, so "be concise" does not erase "go deep for this critique."
Tombstones
Forgetting blocks recapture without exposing the forgotten content in proposals, logs, or proof.
Feedback signals
Route outcomes create weak, bounded observations. Repeated retrieval is not treated as repeated truth confirmation.
The operator workflow
You can inspect and curate the graph from OpenClaw or Telegram with the new /brain graph commands.
/brain graph health
/brain graph dry-run
/brain graph proposals
/brain graph explain <proposal-id>
/brain graph apply <proposal-id>
/brain graph reject <proposal-id>
/brain graph stale
/brain graph clusters
/brain graph tombstonesLow-risk deterministic repairs can be applied. Semantic merges, stale high-authority changes, privacy-sensitive proposals, and scoped exceptions require review.
Public-safe proof
Every mutation is proposal-driven and leaves a redacted trace: run record, proposal record, preconditions, lineage row, edge observation, and proof event. That means the graph can improve without becoming mysterious.
GET /plugins/openclawbrain/graph/health
GET /plugins/openclawbrain/graph/dry-run
GET /plugins/openclawbrain/graph/proposals
POST /plugins/openclawbrain/graph/applyApply-time checks reject stale proposals. Tombstoned and hard-deleted content cannot be revived by merge, split, summarization, proof, proposal, or LLM distillation.
The product difference
| Generic memory | OpenClawBrain |
|---|---|
| Stores or retrieves old context. | Governs memory as evidence with provenance, scope, validity, correction, forgetting, and proof. |
| Can get noisier as it remembers more. | Proposes cleanup so the graph becomes smaller, safer, and more useful. |
| Treats associations as helpful by default. | Separates behavioral association from truth, authority, and privacy. |
| May forget by deleting only the visible note. | Uses tombstones to block unsafe recapture without preserving sensitive content. |