Brain CMS 🧠

A neuroscience-inspired memory architecture for OpenClaw agents. Replaces flat file injection with sparse, semantic, frequency-gated memory loading.

What This Installs

memory/
├── INDEX.md          ← Hippocampus: topic router + cross-links
├── ANCHORS.md        ← Permanent high-significance event store
└── schemas/          ← Domain-specific semantic schemas (you create these)

memory_brain/
├── index_memory.py   ← Embeds schemas into LanceDB vector store
├── query_memory.py   ← Semantic similarity search
├── nrem.py           ← NREM sleep cycle (compression + anchor promotion)
├── rem.py            ← REM sleep cycle (LLM consolidation via Ollama)
└── vectorstore/      ← 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 → 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: 1. Create memory/<topic>.md 2. Add to INDEX.md with triggers + priority + cross-links 3. Re-index: memory_brain/.venv/bin/python3 memory_brain/index_memory.py

Sleep cycles:

# NREM — run on shutdown (~30s, no LLM)
cd ~/.openclaw/workspace/memory_brain && .venv/bin/python3 nrem.py

# REM — 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 — 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)