clawdbrunner/openclaw-graphiti-memory

最近记录的来源活动: 2026年7月22日 12:03

SkillsMP 收录数据更新: 2026年11月12日 12:52

已收集 skills: 1

GitHub 星标: 102
GitHub Forks: 9

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/clawdbrunner/openclaw-graphiti-memory.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# OpenClaw Hybrid Memory: QMD + Graphiti + Shared Files A complete three-layer memory system for [OpenClaw](https://openclaw.ai) multi-agent setups. ## Architecture ``` ┌─────────────────────────────────────────────────────┐ │ Agent Memory │ │ │ │ Layer 1: Private Files (QMD) │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ piper/ │ │ paige/ │ │ dean/ │ ... │ │ │ memory/ │ │ memory/ │ │ memory/ │ │ │ └──────────┘ └──────────┘ └──────────┘ │ │ Per-agent vector search via memory_search │ │ │ │ Layer 2: Shared Files (shared/) │ │ ┌──────────────────────────────────┐ │ │ │ user-profile.md │ │ │ │ agent-roster.md │ │ │ │ infrastructure.md │ │ │ │ graphiti-memory.md │ │ │ └──────────────────────────────────┘ │ │ Symlinked into each agent workspace as shared/ │ │ Indexed by QMD alongside private files │ │ │ │ Layer 3: Shared Knowledge Graph (Graphiti) │ │ ┌──────────────────────────────────┐ │ │ │ openclaw-clawd (write own) │ │ │ │ openclaw-piper (write own) │ │ │ │ user-main (orchestrator) │ │ │ │ system-shared (orchestrator) │ │ │ └──────────────────────────────────┘ │ │ Cross-group search for temporal facts │ └─────────────────────────────────────────────────────┘ ``` ## Three Layers | Layer | What | Best For | Mutability | |-------|------|----------|------------| | **Private Files** | Agent's own `memory/` dir | Private notes, task logs, local state | Agent writes freely | | **Shared Files** | `shared/` dir (symlinked) | Stable reference docs (profiles, roster) | Orchestrator maintains | | **Shared Graph** | Graphiti knowledge graph | Temporal facts, cross-agent knowledge | Agents write to own group | ### When to use which - **"What's the user's email?"** → Shared files (`user-profile.md`) or Graphiti - **"What did I log yesterday?"** → Private files (`memory_search`) - **"What did Piper find about that invoice?"** → Graphiti (cross-group search) - **"Who handles security?"** → Shared files (`agent-roster.md`) - **"When did we change the deployment config?"** → Graphiti (temporal) --- ## Setup ### Prerequisites - [OpenClaw](https://openclaw.ai) installed and configured - Docker (for Graphiti + Neo4j) - An OpenAI API key (for Graphiti embeddings) ### 1. Install Graphiti Stack ```bash git clone https://github.com/clawdbrunner/openclaw-graphiti-memory.git cd openclaw-graphiti-memory # Copy docker-compose to your services directory mkdir -p ~/.openclaw/services/graphiti cp docker-compose.yml ~/.openclaw/services/graphiti/ # Set your OpenAI key export OPENAI_API_KEY="sk-..." # Start the stack cd ~/.openclaw/services/graphiti docker compose up -d ``` ### 2. Configure QMD in OpenClaw Add to `~/.openclaw/openclaw.json` under `agents.defaults`: ```json { "memorySearch": { "enabled": true, "sources": ["memory", "sessions"], "provider": "gemini", "model": "gemini-embedding-001", "sync": { "onSessionStart": true, "watch": true } } } ``` ### 3. Set Up Shared Directory ```bash # Copy scripts to OpenClaw scripts directory mkdir -p ~/.openclaw/scripts cp scripts/*.sh scripts/*.py ~/.openclaw/scripts/ chmod +x ~/.openclaw/scripts/*.sh # Copy shared reference files mkdir -p ~/.openclaw/shared cp shared-files/*.md ~/.openclaw/shared/ # Symlink into each agent's workspace for agent_dir in ~/.openclaw/agents/*/; do agent=$(basename "$agent_dir") [[ "$agent" == "_shared" || "$agent" == "_template" ]] && continue ln -sf ~/.openclaw/shared "$agent_dir/shared" done ``` ### 4. Add Memory Instructions to Agent AGENTS.md Add the shared memory section to each agent's AGENTS.md. See `templates/shared-memory-snippet.md` for a copy-paste template, or use the patch script: ```bash python3 scripts/patch-shared-memory.py ``` ### 5. Seed Shared Context ```bash # Seed user profile scripts/graphiti-log.sh user-main system "System" "User lives in Example City, EST timezone." # Seed agent roster scripts/graphiti-log.sh system-shared system "System" "Agent team: Clawd (orchestrator), Piper (email), Paige (finance)..." ``` --- ## Scripts ### For Agents (`scripts/`) | Script | Purpose | |--------|---------| | `graphiti-search.sh "query" [group_id] [max]` | Search knowledge graph | | `graphiti-log.sh <agent_id> <role> <name> "content"` | Log facts to own group | | `graphiti-context.sh "task" [agent_id]` | Get full context for a task | ### For Setup (`scripts/`) | Script | Purpose | |--------|---------| | `memory-hybrid-search.sh "query"` | Search QMD + Graphiti together | | `graphiti-import-files.py` | Bulk import files into Graphiti | | `graphiti-sync-sessions.py` | Sync session transcripts to Graphiti | | `graphiti-watch-files.py` | Watch files and auto-sync to Graphiti | | `patch-shared-memory.py` | Patch all agent AGENTS.md files | --- ## Graphiti Groups | Group | Owner | Purpose | |-------|-------|---------| | `openclaw-<agent_id>` | Each agent | Agent's own discoveries and decisions | | `user-main` | Orchestrator | User profile, preferences, contacts | | `system-shared` | Orchestrator | Agent roster, infrastructure, active projects | ### Rules 1. Agents write to their **own group only** 2. Agents read **cross-group** (omit `group_id` for global search) 3. Only the orchestrator writes to `user-main` and `system-shared` 4. Shared files in `shared/` are **read-only** for agents — report updates to orchestrator --- ## Docker Compose The included `docker-compose.yml` runs: - **Graphiti API** (port 8001) — REST API for the knowledge graph - **Neo4j** (ports 7474/7687) — Graph database backend Environment variables: - `OPENAI_API_KEY` — Required for embeddings - `MODEL_NAME` — LLM for entity extraction (default: `gpt-4.1-mini`, recommend `gpt-4.1`) --- ## Cost - **QMD:** Free (local, uses Gemini embeddings which are free tier) - **Graphiti:** OpenAI API costs for entity extraction during ingestion only - `gpt-4.1`: ~$2/M input, $8/M output tokens - Searches are free (local Neo4j queries) - Typical cost: < $1/month for a 20-agent setup --- ## License MIT

正在加载仓库 README…

这个仓库中的 skills

软件开发工程师

1 个职业分类 · 已分类 100%

skill

职业分类

描述

更新

hybrid-memory

软件开发工程师

Hybrid memory strategy combining OpenClaw's built-in QMD vector memory with Graphiti temporal knowledge graph. Use for all memory recall requests.

原文语言：英语

2026-07-22

已展示 1 / 1 个已收集 Skill。