OpenClaw Memory Pro System
An AI memory assistant that turns fragmented notes and conversations into searchable long-term memory, auto-distills actionable skills via a closed-loop feedback pipeline, and proactively reminds you.
When to Use
| Goal | Command |
|---|---|
| Store a memory | memory-cli remember "Learned X today" --tag thought -i 0.8 |
| Assembled recall (skills + KG + evidence) | memory-cli recall "X" |
| Deep multi-hop reasoning | memory-cli deep-recall "complex question" |
| Inspiration collision (7 strategies) | memory-cli collide |
| Daily briefing | memory-cli briefing |
| List skills with utility stats | memory-cli skills |
| KG contradiction detection | memory-cli contradictions |
| KG blind spot scan | memory-cli blindspots |
| Thought threads | memory-cli threads |
| Skill feedback | memory-cli skill-feedback <id> success |
When Not to Use
- For ephemeral throwaway messages that don't need persistence.
- For real-time streaming data (this is a batch/on-demand system).
Architecture
Fragments --> [Ingest + Tag] --> Unified Corpus (Memora vectors + MSA documents)
|
+---------+-----------+
v v v
[KG Weave] [Distill] [Collide]
structural compression novelty
_gain _value (1-5)
| | |
+----+----+-----+-----+
| v |
[Skill Proposer] <-- triggered when 2-of-3 scores pass
v
[Skill Registry] <-- utility tracking + feedback loop
(draft -> active -> deprecated)
|
+---------+-----------+
v v v
[Question- [Scheduled [Nebius
Driven Push] Fine-
Recall] Tuning]
|
+-----------+-----------+
v v v
[Skills] [KG Relations] [Evidence] <-- three-layer assembled output
|
v
Use -> Feedback -> utility update -> low-utility auto-rewrite
Subsystems
| Layer | Module | Role |
|---|---|---|
| Corpus | Memora | Primary vector store (nomic-embed-text, JSONL). All content enters here. |
| MSA | Document-level storage for long text (>=100 words) or high importance (>=0.85). LLM-powered multi-hop interleave. | |
| Intelligence | Second Brain | KG weaving, distillation, collision (7 strategies with attention focus + recency weighting). |
| Skill Proposer | Auto-generates draft skills when 2-of-3 scores meet thresholds. | |
| Skill | Skill Registry | Versioned skills with utility tracking, feedback loop, executable action bindings (prompt_template / tool_call / webhook). |
| Training | Chronos | Replay buffer, personality profile generation, training data export. |
Ingestion Routing
- All content -> Memora (always)
- Long text (>=100 words) OR high importance (>=0.85) -> also MSA
- High importance (>=0.85) -> also Chronos
- Always writes daily log file
- Post-remember hooks: KG extraction, access tracking
Recall
Three-layer assembled response with token budget control (default 4000 tokens):
- Skills (score 1.0) — active skills matched by vector similarity, with executable prompts
- KG Relations (score 0.9) — knowledge graph nodes + logical edges
- Evidence (score 0.0-1.0) — Memora snippets + MSA documents
Collision Engine
7 strategies with attention-aware anchor selection:
- RAG-based: Semantic Bridge, Dormant Revival, Temporal Echo, Chronos Cross-Ref, Digest Bridge
- KG-driven: Contradiction-Based, Blind Spot-Based
Before each round, extracts 3-5 focus keywords from recent memories. Anchor selection biased toward current focus topics with recency weighting.
Requirements
- Python 3.9+
- macOS (Apple Silicon) or Linux
- LLM API key: OpenRouter (preferred) or xAI (fallback)
Setup
See setup.md for installation instructions.