anamnesis-hub#

Four-tier memory architecture with automated Dreaming pipeline, three-way synchronization, Active Memory recall, and two-stage auto-extraction.

Overview#

TierLayerTechnologyPurposeL0Runtime Retrievalmemory-core plugin (Ollama bge-m3 → SQLite + sqlite-vec)Real-time semantic + BM25 hybrid searchL0Cloud RecallMemOS Cloud plugin ( optional)Cross-device memory capture and recallL0Active MemoryBuilt-in OpenClaw pluginPre-reply sub-agent memory search + context injectionL1Working Memorymemory/YYYY-MM-DD.md filesDaily summaries, todos, technical notesL2aLong-term IndexMEMORY.md (~90 lines, read-only base)Bidirectional index → ARCHIVE.md, 0% Auto-Extracted pollutionL2bDetailed ArchiveARCHIVE.md (~220 lines)Full records with ← MEMORY.md:XX reverse referencesL2cStructured KBfacts.sqlite (entity/key/value)Precise lookup for IPs, ports, versions; activation/decay tracking

Automated Pipelines#

PipelineScheduleDescriptionDreaming03:00 UTC dailyScan logs → DeepSeek analysis → promote to L2Three-way Sync18:00 / 20:00 / 22:00 CSTCloud ↔ Markdown ↔ Vector alignmentauto-memory v318:30 / 22:30 CSTMemOS Cloud facts → qwen3 filter → Stage1: ARCHIVE.md (SHA-256 dedup) → Stage2: MEMORY.md (summary)session-extract22:00 CSTScan session JSONL → MemOS extractor/reranker → .learnings/ + memory/YYYY-MM-DD.md (two-pass)

Memory Flow (Unified Daily Pipeline)#

text

agent_end → MemOS Cloud (cloud extraction)
              ↓
    ┌─ 18:00 sync-pull (Cloud → local cache) ─┐
    │  18:30 auto-memory.py                    │
    │  → Stage 1: ARCHIVE.md (SHA-256 dedup)   │
    │  → Stage 2: MEMORY.md (one-line summary) │
    │                                           │
    │  20:00 sync-push + reindex                │
    │  22:00 session-extract.py                 │
    │  → Pass 1: .learnings/ERRORS.md           │
    │  → Pass 2: memory/YYYY-MM-DD.md           │
    │  → archive analyzed → trash               │
    └───────────────────────────────────────────┘

before_agent_start → MemOS Cloud recall
                  + Active Memory sub-agent
                  + memory-core (bge-m3 + BM25)
              ↓
        Layered context injected before reply

Scripts#

ScriptPurposeauto_memory_extract.pyv3 two-stage pipeline: ARCHIVE.md archive → MEMORY.md summarysession-extract.pyScan session JSONL → .learnings/ + memory/ (two-pass)seed-facts-db.pyInitialize facts.sqlite from ARCHIVE.mdfacts_activation.pyHebbian activation + daily decay + Hot/Warm/Cooldaily-memory-pipeline.sh6-stage unified daily pipelinewrite_file.pyCross-platform text writer (UTF-8/BOM/CRLF)auto-setup.shOne-command Ollama + memory-core + MemOS Cloud setup

Setup#

One-command auto-setup#

bash

bash scripts/auto-setup.sh

Options#

bash

bash scripts/auto-setup.sh --skip-ollama   # Skip Ollama install
bash scripts/auto-setup.sh --skip-memos    # Skip MemOS Cloud
bash scripts/auto-setup.sh --dry-run       # Preview only

Manual setup#

See references/setup-guide.md for step-by-step manual configuration.

When to Use#

• Setting up OpenClaw memory for the first time
• Configuring memory-core plugin with local Ollama embedding
• Installing MemOS Cloud plugin for cross-device sync
• Enabling Active Memory for pre-reply context injection
• Setting up auto-memory pipeline for MEMORY.md curation
• Installing cross-platform-writer for cross-OS file compatibility
• Configuring automatic Dreaming and promotion pipelines
• Setting up three-way sync between cloud, files, and vector DB

Plugin Conflicts#

❌ subconscious-personality-guardian ↔ memory-core#

Incompatible. Both use the same OpenClaw memory slot.

Fix: Disable in openclaw.json:

json

{
  "plugins": {
    "disabled": ["subconscious-personality-guardian"],
    "deny": ["subconscious-personality-guardian"]
  }
}

✅ memory-core + MemOS Cloud#

Compatible — designed to work in layers.

text

User message → MemOS Cloud (static facts) → memory-core (recent context)

✅ Active Memory + MemOS Cloud + memory-core#

Compatible — triple-layer recall.

text

User message
  → Active Memory sub-agent (searches all memory stores)
  → MemOS Cloud (injects long-term facts & preferences)
  → memory-core (semantic + BM25 hybrid retrieval)
  → Agent receives layered context

✅ auto-memory + MemOS Cloud + memos-extractor#

Designed to work together. MemOS captures at agent_end, syncs to local files, then auto-memory reads those files for MEMORY.md curation. v1.10 adds a second channel: memos-extractor-0.6b API (MemOS self-developed model) returns structured facts + preferences, cross-validated against qwen3 output.

Components#

1. Memory Plugins (L0)#

See references/architecture.md for full configuration.

2. Memory Files (L1 + L2)#

text

~/.openclaw/workspace/
├── memory/
│   ├── YYYY-MM-DD.md          # Daily working memory (auto-indexed)
│   ├── MEMORY_INDEX.md        # Vector BM25 cluster summaries
│   └── .sync-*.json           # Sync state files
├── MEMORY.md                  # Long-term memory index (~90 lines)
├── ARCHIVE.md                 # Detailed archive (~220 lines, ← bidirectional refs)
├── AGENTS.md                  # Runtime context + memory rules
└── user_workspace/
    ├── memos-cloud-cache/         # Cloud-pulled memory (isolated from index)
    │   └── memos-cloud-*.md
    ├── scripts/
    │   └── sync-*.py              # Sync scripts (pull/push/vector)
    └── skills/
        └── anamnesis-hub/   # Installed from ClawHub
            └── scripts/           # Pipeline scripts (auto-memory, session-extract, etc.)

3. Pipeline Scripts (in anamnesis-hub/scripts/)#

• auto_memory_extract.py — v3 two-stage pipeline (ARCHIVE.md → MEMORY.md)
• session-extract.py — Session JSONL scan → .learnings/ + memory/
• seed-facts-db.py — Initialize facts.sqlite from ARCHIVE.md
• facts_activation.py — Hebbian activation + daily decay
• daily-memory-pipeline.sh — 6-stage unified daily pipeline
• write_file.py — Cross-platform text file writer
• auto-setup.sh — One-command Ollama + memory-core + MemOS Cloud setup

See references/INDEX.md for full documentation index.

File Reference#

👉 完整文档请查阅 references/INDEX.md

文档说明references/INDEX.md📌 文档入口索引（快速定位）references/architecture.md四层架构设计、插件配置、协同流程references/setup-guide.md环境配置、手动安装步骤、cron 设置references/memory-directory.mdmemory/ 目录结构、状态文件、LCM 机制references/sync-api.mdMemOS Cloud API 参考references/scripts-reference.md所有脚本统一说明（用途、cron、依赖）references/pipeline-stages.md日终管线 6 阶段详解references/candidates-review.mdP3 记忆候选审核机制references/upgrade-reset.md升级路径、重置流程、卸载步骤