Skip to content
BASS

Memory System

Memory System

Section titled “Memory System”

Principle: Disk is Source of Truth

Section titled “Principle: Disk is Source of Truth”

All agent knowledge lives on disk. Write outcomes to files immediately. Memory in context is ephemeral — disk is canonical.

Directory Structure

Section titled “Directory Structure”
/home/chrisr6/.openclaw/workspace/
├── MEMORY.md           # Long-term curated essentials (agent reads on startup)
├── HANDOVER.md         # Machine-readable session handover (overwritten each session)
├── HEARTBEAT.md        # Current Now/Next/Blockers (updated each heartbeat)
├── AGENTS.md           # Delegation + spawning rules
├── IDENTITY.md         # Tila's identity and role
├── SOUL.md             # Core operating principles + checkpoint protocol
├── memory/
│   ├── daily/          # Append-only session logs (YYYY-MM-DD.md)
│   ├── decisions/      # CEO/Tila decisions by month (YYYY-MM.md)
│   ├── projects/       # Per-project status files
│   ├── ops/            # Operational reference docs
│   ├── research/       # Research outputs by topic
│   └── checkpoints/    # Sub-agent result files (subagent-<label>.md)

File Types

Section titled “File Types”

Ephemeral (session-scoped)

Section titled “Ephemeral (session-scoped)”
  • HANDOVER.md — Overwritten on every checkpoint.
  • HEARTBEAT.md — Updated every 30 minutes.

Append-Only

Section titled “Append-Only”
  • memory/daily/YYYY-MM-DD.md — Session summaries.

Canonical Reference

Section titled “Canonical Reference”
  • MEMORY.md — Curated essentials.
  • AGENTS.md — Delegation rules, routing, agent fleet table.
  • memory/ops/*.md — Operational procedures, model catalog, corrections log.
  • memory/projects/*.md — Per-project status.

Memory Protocol

Section titled “Memory Protocol”

On session startup, agents read:

  1. HANDOVER.md (last session state)
  2. MEMORY.md (curated long-term memory)
  3. HEARTBEAT.md (current status)
  4. Project-specific files (FRANKEL.md, etc.)

Vector Backend

Section titled “Vector Backend”
  • Engine: Local embeddings (embeddinggemma-300m)
  • State: 14 files, 72 chunks
  • Use memory_search for deep lookups across transcript history

See Also

Section titled “See Also”