javimosch Brain CMS π§
A neuroscience-inspired memory architecture for OpenClaw agents. Replaces flat file injection with sparse, semantic, frequency-gated memory loading.
What This Installs
memory/
βββ INDEX.md β Hippocampus: topic router + cross-links
βββ ANCHORS.md β Permanent high-significance event store
βββ schemas/ β Domain-specific semantic schemas (you create these)
memory_brain/
βββ index_memory.py β Embeds schemas into LanceDB vector store
βββ query_memory.py β Semantic similarity search
βββ nrem.py β NREM sleep cycle (compression + anchor promotion)
βββ rem.py β REM sleep cycle (LLM consolidation via Ollama)
βββ vectorstore/ β LanceDB database (auto-created)
Setup (one-time)
# 1. Run the installer
python3 ~/.openclaw/workspace/skills/brain-cms/install.py
# 2. Index your schemas
cd ~/.openclaw/workspace/memory_brain
.venv/bin/python3 index_memory.py
# 3. Test retrieval
.venv/bin/python3 query_memory.py "your topic here" --sources-only
How It Works
Boot sequence: Load MEMORY.md (lean core) + today's daily log. Nothing else.
When a topic appears: Read memory/INDEX.md β load only the relevant schemas (spreading activation). Check memory/ANCHORS.md for high-significance events.
For ambiguous topics: Run semantic search:
memory_brain/.venv/bin/python3 memory_brain/query_memory.py "message text" --sources-only
Auto-schema creation: When a new significant project or domain appears:
- Create
memory/<topic>.md - Add to INDEX.md with triggers + priority + cross-links
- Re-index:
memory_brain/.venv/bin/python3 memory_brain/index_memory.py
Sleep cycles:
# NREM β run on shutdown (~30s, no LLM)
cd ~/.openclaw/workspace/memory_brain && .venv/bin/python3 nrem.py
# REM β run weekly (2-5 min, uses local llama3.2:3b, free)
cd ~/.openclaw/workspace/memory_brain && .venv/bin/python3 rem.py
Memory Layers (CMS)
| Layer | Files | When loaded | Purpose |
|---|---|---|---|
| Working | MEMORY.md + today log |
Every session | Core context |
| Episodic | memory/YYYY-MM-DD.md |
Session boot | Recent events |
| Semantic | memory/*.md schemas |
On trigger | Domain knowledge |
| Anchors | memory/ANCHORS.md |
On CRITICAL topics | Permanent ground truth |
| Vector | memory_brain/vectorstore/ |
On demand | Semantic search |
Tagging Anchors
In any daily log, tag high-significance events:
[ANCHOR] Major demo success β full pipeline working end-to-end
NREM auto-promotes these to ANCHORS.md on next shutdown.
Token Savings
Typical MEMORY.md: 150-300 lines injected every session. With Brain CMS: ~50-line core + schemas loaded only when relevant. Estimated savings: 40-60% reduction in context tokens per session.
Requirements
- Python 3.10+
- Ollama (for embeddings + REM consolidation)
- 500MB+ storage for vector store and models
lancedb,numpy,pyarrow,requests(auto-installed)