MemCore 记忆核心#

为 OpenClaw 打造的五层自适应记忆系统，自进化。替代全量 MEMORY.md 注入 — 节省 92% token。

Five-tier adaptive memory system with self-evolution for OpenClaw. Drop-in replacement for full MEMORY.md injection — 92% token savings.

v2.1.0 (proactive-agent enhanced) ⭐#

• WAL Protocol: cmd_index now parses SESSION-STATE.md as high-priority traces
• Working Buffer: SESSION-STATE.md + working-buffer.md searchable in World Model tier
• VFM Scoring: 4-dimension trace evaluation (frequency, failure reduction, user burden, self cost)
• New command: python3 scripts/memcore/cli.py vfm — VFM score analysis

Quick Commands#

bash

# Startup: generate brief (≤500 tokens, replaces full MEMORY.md)
python3 scripts/memcore/cli.py brief

# Search: 5-tier adaptive fallback
python3 scripts/memcore/cli.py search "<query>" -n 5

# Daily maintenance (index new logs → induce patterns → decay → refresh brief)
python3 scripts/memcore/cli.py run-all

# VFM score analysis
python3 scripts/memcore/cli.py vfm

# Stats overview
python3 scripts/memcore/cli.py stats

# Feedback: log whether retrieved memory was useful
python3 scripts/memcore/cli.py feedback-log <trace_id> used|skipped|good|bad

Startup Integration#

On each session start, replace full MEMORY.md loading with:

text

1. Read SOUL.md → USER.md → MEMORY_BRIEF.md (auto-generated) → .anatomy.md → yesterday+today memory
2. MEMORY.md loaded on-demand only via: python3 scripts/memcore/cli.py search "<keywords>"

The brief generator picks top 5 active patterns, top 3 recent lessons, and current taskboard items — all under 500 tokens.

散会 Integration#

Add step 3 to 散会 flow:

bash

python3 scripts/memcore/cli.py index && \
python3 scripts/memcore/cli.py induce && \
python3 scripts/memcore/cli.py feedback

This auto-indexes today's log, induces new patterns, and decays stale traces.

Health Check Integration#

When running daily health checks (09:00), append MemCore diagnostic:

bash

# Get system stats
python3 scripts/memcore/cli.py stats

# If anomalies found, search historical similar faults
python3 scripts/memcore/cli.py search "<anomaly keywords>" -n 3

Safety rule: Report only. Never auto-fix. Let human decide.

Search & Feedback Flow#

After every memory_search call, log feedback:

bash

# If the retrieved memory was helpful:
python3 scripts/memcore/cli.py feedback-log <trace_id> used

# If irrelevant:
python3 scripts/memcore/cli.py feedback-log <trace_id> skipped

This trains the retrieval system — high-value traces rise, stale ones decay.

Cron Setup#

Two 12-hour maintenance jobs (systemEvent, main session, wakeMode: next-heartbeat):

Time (CST)Cron expressionAction04:000 4 * * *index + induce + feedback + brief16:000 16 * * *same

These are SQLite-only, no model calls, near-zero resource cost.

Rollback#

bash

cp memcore_backup_<date>/AGENTS.md ~/.openclaw/workspace/
cp memcore_backup_<date>/MEMORY.md ~/.openclaw/workspace/

Original MEMORY.md is never modified. All MemCore data lives in separate SQLite files under ~/.openclaw/.

Detailed Reference#

• Architecture: See references/architecture.md for L1-L4 model, 5-tier retrieval details, file structure, and CLI reference.
• EXO1 Upgrade: See references/upgrade-guide.md for Syncthing-based deployment to EXO1.