Memory Dream — 记忆分类与整合系统

Inspired by Claude Code's Dream memory consolidation system. Adapted for OpenClaw's file-based memory architecture.

Memory Classification — 四类记忆

Every memory entry belongs to one of four types. Tag each entry when writing.

Writing Format

Every feedback/project memory entry should follow:

**Rule/Fact:** [the thing itself]
**Why:** [reason — past incident, user preference, or constraint]
**How to apply:** [when/where this kicks in]

Key Principles

1. Record success too — Only recording failures makes the agent overly cautious. "This approach worked well + why" is as valuable as "this broke + why"
2. Include Why — Without Why, the agent can't judge edge cases and blindly follows rules
3. Absolute dates only — Write "2026-03-31" not "today" or "yesterday". Memories must be interpretable after time passes
4. Verify before trusting — "Memory says X exists" ≠ "X exists now". Check current state before acting on recalled memories

What NOT to Store

• Code patterns, architecture, file structure (derivable from code)
• Git history (git log is authoritative)
• Debugging solutions (the fix is in the code, commit message has context)
• Ephemeral task details or current conversation context
• Content already in CLAUDE.md / AGENTS.md

Even if the user asks to save something from the exclusion list, ask what was surprising or non-obvious — that's the part worth keeping.

MEMORY.md Size Control

Hard limit: 200 lines / 25KB. When exceeded, run a Dream cycle.

Check with: wc -l MEMORY.md && wc -c MEMORY.md

Dream Cycle — 记忆整合流程

A periodic consolidation pass over memory files. Run during heartbeats or when MEMORY.md exceeds limits.

Trigger Gates (all three must pass)

1. Time gate: ≥ 3 days since last Dream
2. Session gate: ≥ 3 days of new daily notes since last Dream
3. Lock gate: No other Dream in progress

Track state in memory/heartbeat-state.json:

{
  "lastDreamAt": "2026-03-31T15:40:00Z",
  "lastDreamResult": "pruned from 269 to 85 lines"
}

Four Phases

Phase 1 — Orient

• Read MEMORY.md — understand current long-term memory landscape
• Scan recent memory/YYYY-MM-DD.md files since last Dream
• Note current MEMORY.md line count and byte size

Phase 2 — Gather

• Extract entries worth keeping long-term from daily notes
• Classify each as user / feedback / project / reference
• Priority: important decisions, then lessons learned, then discoveries, then people info
• Collect successes too, not just failures

Phase 3 — Consolidate

• Write or update MEMORY.md entries in appropriate sections
• Convert any relative dates to absolute dates
• Delete contradicted facts (new info overrides old)
• Merge near-duplicate entries
• For feedback type: ensure each has Why + How to apply

Phase 4 — Prune

• Enforce ≤ 200 lines / 25KB limit
• Remove: completed projects, resolved issues, stale preferences
• Replace verbose details with pointers: "详见 daily notes 2026-03-31"
• Check for and resolve contradictions between entries

After Dream

• Update lastDreamAt in heartbeat-state.json
• Log what was done in today's daily notes: lines before → after, what was added/removed

Setup Guide — First-Time Configuration

To add this system to a new OpenClaw agent:

1. Add the memory classification table and Dream procedure to AGENTS.md
2. Add size limit header to MEMORY.md:

   > 📏 限制：≤ 200 行 / 25KB | 上次整理：YYYY-MM-DD
   

3. Add lastDreamAt field to memory/heartbeat-state.json
4. Add Dream trigger check to HEARTBEAT.md:

   - [ ] **记忆整理（Dream）**：检查是否满足三重门控，满足则执行四阶段整理
   

5. Run an initial Dream cycle to establish baseline

Memory Drift Warning

Before acting on a recalled memory:

• If memory names a file path → check the file exists
• If memory names a function or config → grep for it
• If memory is a state snapshot (activity log, architecture) → prefer git log or reading current code
• If memory conflicts with current reality → trust what you see now, update the memory