Dream — Nightly Memory Distillation & Self-Evolution

现状

Phase 1（✅ 已实现）： OpenClaw Hook hippocampus 监听每条消息，实时写入 memory/logs/ Phase 2（✅ 已实现）： dream.py v3.3 定时蒸馏，新增 Distillation Agent 做真正的推理

完整链路

用户对话
   ↓
OpenClaw Hook: message:preprocessed
   ↓
丘脑过滤（Thalamus）→ 杏仁核标记（Amygdala）→ 海马体存储（memory/logs/）
   ↓
cron 触发（早7点/晚10点）
dream.py v3.3
   ↓
读取 hippocampus 日志 + recall store
   ↓
杏仁核标记融合 → Auditor审计 → 分析皮层模式识别 → 前额叶蒸馏规划 → 归档区 → 真相文件写回 → 梦境报告

脑区协同架构

① 丘脑（Thalamus）— 注意力门控

过滤纯问候/简单确认，只记录有意义的事件 标记类型：event / decision / correction / completed / insight / error

② 杏仁核（Amygdala）— 情绪标记

correction/error/decision/completed/insight 携带 HIGH 权重，优先蒸馏

③ 海马体（Hippocampus）— 原始存储

Phase 1：OpenClaw hook 实时追加到 memory/logs/YYYY/MM/YYYY-MM-DD.md Phase 2：dream.py 读取日志，融合 recall store，一并蒸馏

④ 前额叶（Prefrontal Cortex）— 蒸馏规划 + 执行控制

人脑中前额叶负责计划、推理、抑制冲动。核心功能是「在适当的时候做适当的事」。 Dream 中：

• 规划：cron 触发时决定今天沉淀什么、遗忘什么
• 抑制：断路器规则——3轮内无法完成则终止，防止死循环
• 调度：协调分析皮层和归档区的工作顺序

⑤ 蓝斑核（Locus Coeruleus）— 警觉与新鲜度信号

人脑中蓝斑核释放去甲肾上腺素，调节警觉和注意力，对新奇事物敏感。 Dream 中：health metrics 里的 freshness 分数——最近被提及的记忆权重更高。 公式：freshness = entries_referenced_last_30_days / total_entries

Phase 1 — 丘脑 + 杏仁核 + 海马体（始终活跃）

对话进行时，三个脑区协同工作：

1. 丘脑过滤：「这条要记吗？」→ 重要事件才写日志
2. 杏仁核标记：correction / error 携带 [HIGH] 权重
3. 海马体存储：追加到 memory/logs/YYYY/MM/YYYY-MM-DD.md

Phase 2 — 前额叶 + 分析皮层 + 蓝斑核（cron 或手动触发）

cron 触发时：

1. 前额叶规划：读取今日日志 + .learnings/ + 现有 topic 文件，制定蒸馏计划
2. 分析皮层扫描：失败模式 / 重复错误 / 用户纠正规律 → 提炼 Pattern
3. 蓝斑核评分：计算 freshness / coverage / coherence → 更新 health metrics
4. 归档区执行：90天未引用且低价值的条目归档
5. 前额叶输出：梦境报告（生长指标 + 洞察 + 明日重点）

Integration with self-improving-agent

Dream reads and promotes entries from the self-improving-agent skill:

.learnings/
├── LEARNINGS.md        # Corrections, insights, best practices
├── ERRORS.md          # Command failures, integration errors
└── FEATURE_REQUESTS.md # User-requested capabilities

Distillation logic:

• Corrections (category: correction) → promote to feedback topic
• Recurring errors (3+ occurrences) → extract to Pattern Library
• Errors with fixes → promote to relevant topic
• Feature requests → create project topic if validated
• Best practices → promote to relevant topic
• Mark promoted entries as status: promoted

Pattern Library (from eo-ability-dream)

Patterns are reusable response templates extracted from recurring learnings:

memory/patterns/
└── p-xxx.md           # Pattern files with trigger + response

Pattern format:

---
name: pattern名称
trigger: 什么情况下触发
response: 如何响应
examples: [案例1, 案例2]
created: YYYY-MM-DD
updated: YYYY-MM-DD
---

Extraction criteria:

• Same error occurs ≥3 times across logs
• User-corrected mistake ≥2 times
• Non-obvious fix discovered through debugging

Memory Taxonomy

Four memory types:

• user — User preferences, role, goals, communication style
• feedback — User corrections and confirmations
• project — Project context, decisions, deadlines, tools in use
• reference — External systems, links, credentials locations

Directory Structure

memory/
├── logs/
│   └── YYYY/MM/YYYY-MM-DD.md   # Daily append-only logs
├── topics/                      # Distilled topic memories
│   ├── user_xxx.md
│   ├── feedback_xxx.md
│   ├── project_xxx.md
│   └── reference_xxx.md
├── patterns/                    # Pattern Library (from eo-ability-dream)
│   └── p-xxx.md               # Reusable patterns
├── episodes/                    # Project narratives (from openclaw-auto-dream)
├── procedures.md               # Workflow preferences
├── archive.md                  # Compressed old entries
├── dream-log.md               # Dream cycle reports
├── index.json                 # Entry metadata (v3.0 schema)
└── MEMORY.md                   # INDEX only — points to topics

.learnings/                     # self-improving-agent
├── LEARNINGS.md
├── ERRORS.md
└── FEATURE_REQUESTS.md

Health Score (from openclaw-auto-dream)

Five metrics measured each dream cycle:

Health Score = weighted sum × 100 (0-100)

Importance Scoring (from openclaw-auto-dream)

importance = base_weight × recency_factor × reference_boost

• base_weight: default 1.0, 🔥HIGH = 2.0, ⚠️PERMANENT = 1.0 (skip formula)
• recency_factor: max(0.1, 1.0 - days/180)
• reference_boost: max(1.0, log2(refs + 1))

Archive conditions (all must be true):

• days_since_referenced > 90
• importance < 0.3
• NOT ⚠️PERMANENT, NOT 📌PIN, NOT in episodes/

Optional: SwarmRecall Cloud Tier

Only active if SWARMRECALL_API_KEY environment variable is set.

SwarmRecall provides advanced cloud-side operations for large memory stores:

• Tier 1 (server-side): deduplication clustering, importance decay, stale pruning
• Tier 2 (agent-driven): contradiction resolution, session summarization, knowledge graph enrichment

Setup

1. Get an API key: POST /api/v1/register with { "name": "<agent-name>" }
2. Set environment variable: SWARMRECALL_API_KEY=<your-key>
3. SwarmRecall is now active — Tier 1 runs automatically on configured interval

SwarmRecall Endpoints Used

• POST /api/v1/dream — start dream cycle
• POST /api/v1/dream/execute — run server-side ops (decay, prune)
• GET /api/v1/dream/candidates/duplicates — find duplicate clusters
• GET /api/v1/dream/candidates/contradictions — find conflicting memories
• GET /api/v1/dream/candidates/stale — find prunable memories
• PATCH /api/v1/dream/:id — report dream results

Privacy

• Data transmitted over HTTPS to swarmrecall-api.onrender.com
• Only your agent's existing memory data is sent — no new external data collected
• Archived memories are soft-deleted (recoverable)
• Tenant-isolated by owner ID and agent ID

MEMORY.md Index Format

MEMORY.md is an INDEX ONLY — one line per topic:

# MEMORY.md — Long-Term Memory Index

_Last updated: YYYY-MM-DD_

## Topics
- [Title](topics/file.md) — one-line description
- [Title](topics/file.md) — one-line description

## Patterns
- [pattern-name](patterns/p-xxx.md) — trigger summary

## Stats
- Total entries: N | Health: N/100 | Streak: N dreams

Daily Log Format

Append to memory/logs/YYYY/MM/YYYY-MM-DD.md after each significant event:

## HH:MM

- [type] content description
- completed: task name
- decision: what was decided
- correction: user feedback received
- insight: what was learned

Types: event | decision | correction | completed | insight | error | question

Cron Configuration

Two cron jobs:

• dream-morning: 7:00 AM
• dream-night: 10:00 PM
• --no-deliver (writes to file only)

Cron message is updated daily via update-cron-date.py script which injects current date.

Dream Distillation Steps (v3.3)

When cron triggers:

1. Read memory/logs/{date}.md
2. Read .learnings/LEARNINGS.md, .learnings/ERRORS.md, .learnings/FEATURE_REQUESTS.md
3. Read MEMORY.md, topic files, index.json, procedures.md for context
4. Snapshot BEFORE: count entries, decisions, lessons, procedures
5. Analyze failures:
   • Count occurrences of each error in ERRORS.md
   • Identify recurring patterns (≥3 occurrences)
   • Extract root cause
6. Distillation Agent（新增）: Run sub-agent on raw entries + learnings + health metrics → produce:
   • 3-5 genuine insights ("I learned that...")
   • 1-3 tomorrow action items
   • 0-3 topic files to write to memory/topics/
   • Health metric interpretation
7. Write topic files (from Distillation Agent output)
8. Update truth files (user_state.md, pending.md)
9. Update index.json entry metadata (id, importance, tags, related)
10. Compute health metrics → update index.json stats
11. Archive eligible entries → append to archive.md
12. Update MEMORY.md index (max 200 lines)
13. Snapshot AFTER: calculate deltas
14. Write dream report to memory/dreams/{date}.md and dream-log.md
15. Regenerate dashboard.html if exists
16. [Optional SwarmRecall]: If SWARMRECALL_API_KEY is set:
    • Start dream cycle: POST /api/v1/dream
    • Run server-side ops: POST /api/v1/dream/execute
    • Fetch duplicate clusters → merge redundant memories
    • Fetch contradictions → resolve conflicting entries
    • Update dream cycle with results

Dream Report Format (v3.3)

# 🌙 Dream Report — {date}

## System Status
- Raw entries: N | recall store: N | Health: {score}/100

## Health Insights
- {insight based on coherence/freshness metrics}

## Insights ("I Learned")
- {genuine insight 1}
- {genuine insight 2}

## Tomorrow's Focus
- {actionable item 1}
- {actionable item 2}

## Topic Files Written
- {filename}: {title}

## Analysis
- Recurring errors found: {list}
- Root causes identified: {analysis}

## Patterns Updated
- {pattern_name}: {change}

First Dream (post-install)

After initial setup, run a full scan of ALL existing daily logs (not just last 3 days). Compare before/after metrics and send user a First Dream Report showing:

• How many logs scanned, entries extracted
• Before → After table
• Personalized reflection based on actual log content

User Prompts

• "dream report" / "梦境报告" → read and display latest dream report
• "dream" / "做梦" → run distillation now
• "/dream status" → show health score, pattern count, streak, stale threads

Scripts

• dream.py — Phase 2 蒸馏脚本（v3.3，Distillation Agent）
• update-cron-date.py — 每日 cron 日期注入

Phase 1 启用（hippocampus hook）

Hook 目录： ~/.openclaw/hooks/hippocampus/ 已配置： openclaw.json 中 hooks.internal.entries.hippocampus: enabled: true

功能： 监听 message:preprocessed 事件，自动记录对话到 memory/logs/YYYY/MM/YYYY-MM-DD.md

丘脑过滤规则：

• 纯问候 / 简单确认（<20字）不记录
• 高权重标记：correction / error / decision / completed / insight

重启 gateway 后生效：

schtasks /run /tn "OpenClaw Gateway"

技能整合说明

Phase 1 + Phase 2 完整闭环（已实现）

• Phase 1：~/.openclaw/hooks/hippocampus/ — OpenClaw hook 实时记录
• Phase 2：dream.py v3.3 — Distillation Agent 做真正的推理，取代"444个候选topic"的列表

已整合

• self-improving-agent → dream 的日志层（.learnings/），无需单独使用

其他AI自我进化技能定位

• recursive-self-improvement：代码级自修复系统（错误检测→根因分析→修复→测试→验证），与 dream 是不同维度，可共存
• dream = 记忆蒸馏 + Pattern提取 + 健康评分
• recursive-self-improvement = 代码错误自动修复