sleep-consolidation（睡眠巩固）

Agent Sleep Consolidation — v2

Biologically-grounded memory consolidation for AI agents, based on:

• Eichenlaub et al. 2020 (Cell Reports): Neural replay in human motor cortex — waking rest consolidates too; NREM1 shows compressed (fast) AND dilated (slow) dual-track replay; replay prioritizes weakly-learned items
• Walker, "Why We Sleep": NREM hippocampus→cortex transfer, synaptic pruning (~20%), REM creative synthesis
• OpenClaw memory architecture: Markdown-as-truth, two-layer files, pre-compaction flush, hybrid BM25+vector recall

See references/workspace_layout.md for full file format examples. See references/memory_schema.md for Markdown conventions and type system.

Memory workspace layout

~/.agent_workspace/
├── MEMORY.md                  ← long-term: durable facts, preferences, decisions
├── memory/
│   └── YYYY-MM-DD.md          ← daily log (append-only; load today + yesterday)
└── bank/
    ├── entities/              ← per-person or per-project pages
    └── concepts/              ← per-topic deep-dives

Core rule: The agent only "remembers" what gets written to disk. Never keep important things in RAM.

Three consolidation modes

Mode 1 — Micro-rest (waking replay)

Quick append to today's daily log. No LLM call — pure write. Based on the paper's finding that waking replay occurs during rest blocks immediately after learning.

python scripts/micro_rest.py \
  --note "User prefers TypeScript; rejected Python suggestion" \
  --type O \
  --workspace ~/.agent_workspace

# Pre-compaction flush: extract worth-retaining from raw context
python scripts/micro_rest.py \
  --flush \
  --context-dump "$(cat session_log.txt)" \
  --workspace ~/.agent_workspace

Memory type prefixes (W/B/O/S system from OpenClaw research):

Each entry is appended to memory/YYYY-MM-DD.md under a ## Retain section.

Mode 2 — NREM consolidation

Two-track processing based on the paper's temporal replay findings:

• Fast track (compressed replay, ~0.1x duration): high-confidence facts → MEMORY.md
• Slow track (dilated replay, 1.5–2x duration): weakly-learned or uncertain items the paper shows replay *prioritizes* these → bank/ for deeper storage

python scripts/sleep_cycle.py --workspace ~/.agent_workspace --phase nrem

The script reads today's daily log + current MEMORY.md, then calls Claude:

NREM system prompt

You simulate NREM deep-sleep memory consolidation for an AI agent.

Two-track processing:

FAST TRACK (compressed replay — high-confidence, clear facts):
  - Distill into single-sentence durable memories for MEMORY.md
  - Prune ~20% as redundant (synaptic homeostasis: sleep removes weak connections)
  - Use prefixes: W (world fact), B (experience), O(c=N) (opinion + confidence), S (summary)

SLOW TRACK (dilated replay — weakly-learned, uncertain, needing reinforcement):
  - Expand on low-confidence or partial-knowledge items
  - Write to bank/ as entity or concept pages
  - These are what human NREM prioritizes (Schapiro et al. 2018)

Respond ONLY with valid JSON:
{
  "memory_md_updates": [
    {
      "action": "add|update|remove",
      "type": "W|B|O|S",
      "confidence": 0.0,
      "text": "single sentence",
      "replaces": "old text if action=update"
    }
  ],
  "bank_updates": [
    {
      "slug": "topic-slug",
      "kind": "entity|concept",
      "section": "## Section heading",
      "content": "markdown block"
    }
  ],
  "pruned_count": 0,
  "weakly_learned": ["items for slow-track attention"],
  "open_questions": ["gaps for next session"]
}

Mode 3 — REM synthesis

Creative cross-domain connections. The dreaming brain ignores conventional logic.

python scripts/sleep_cycle.py --workspace ~/.agent_workspace --phase rem

REM system prompt

You simulate REM dream-sleep creative synthesis for an AI agent.
Find non-obvious connections. Update beliefs where evidence shifted.

Tasks:
1. Cross-domain links between today's memories and MEMORY.md
2. Update O(c=N) confidence values where evidence has shifted
3. Generate concrete next-session actions
4. Propose new bank/ pages if a topic recurred 3+ times this session

Respond ONLY with valid JSON:
{
  "aha_moments": ["string"],
  "creative_connections": [{"link": "A ↔ B", "why": "string"}],
  "confidence_updates": [
    {"item": "text excerpt", "old_c": 0.0, "new_c": 0.0, "reason": "string"}
  ],
  "new_bank_suggestions": [
    {"slug": "string", "kind": "entity|concept", "seed_content": "markdown"}
  ],
  "next_session_actions": ["string"]
}

Pre-compaction flush

Mirrors OpenClaw's memoryFlush mechanism. Triggered when context approaches token limit. Runs silently (no user output).

Configure the threshold (equivalent to OpenClaw's softThresholdTokens):

# When session token estimate > (context_window - 20000 - 4000):
python scripts/micro_rest.py --flush --context-dump "..." --workspace ~/.agent_workspace

The --flush flag calls Claude to extract durable memories from the raw context dump, then writes them to the daily log. One flush per compaction cycle.

Loading memory at session start

from scripts.load_memory import get_context

system_addendum = get_context(
    workspace="~/.agent_workspace",
    max_memory_chars=3000,   # MEMORY.md snippet (fits ~2k tokens)
    load_yesterday=True      # also load yesterday's daily log
)
# Inject system_addendum into system prompt

Loaded content:

1. Full MEMORY.md (trimmed to max_memory_chars)
2. Today's daily log (if exists)
3. Yesterday's daily log (if load_yesterday=True)

Full session workflow

# Session start: inject memory into system prompt (in Python)

# During session: micro-rest after significant exchanges
python scripts/micro_rest.py --note "..." --type W --workspace ~/.agent_workspace

# If context > 80% full: flush before compaction
python scripts/micro_rest.py --flush --context-dump "..." --workspace ~/.agent_workspace

# Session end: full sleep cycle
python scripts/sleep_cycle.py --phase both --workspace ~/.agent_workspace

Troubleshooting