memvault
Production-ready long-term memory server for AI agents with Ebbinghaus decay and strength-weighted retrieval. Use when you need persistent memory across agent sessions, memory decay (forgetting curve), memory statistics, or multi-agent memory tracking. Triggers: long-term memory, remember, recall, memory decay, Ebbinghaus, agent memory, memvault. Requires Docker.
- License 许可
- MIT
- 最近更新
- 2026-06-10
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- memvault
- description
- Production-ready long-term memory server for AI agents with Ebbinghaus decay and strength-weighted retrieval. Use when you need persistent memory across agent sessions, memory decay (forgetting curve), memory statistics, or multi-agent memory tracking. Triggers: long-term memory, remember, recall, memory decay, Ebbinghaus, agent memory, memvault. Requires Docker.
03 SKILL.md 正文
MemVault — Long-term Memory for AI Agents
Quick Start
# Install (one command — Docker handles everything)
bash scripts/install.sh
# Verify
memvault health
Usage
# Store a memory
memvault memorize-text <user_id> "<message>" "[reply]"
# Retrieve memories (strength-weighted)
memvault retrieve <user_id> "<query>"
# Run daily decay (memories fade like human memory)
memvault decay <user_id>
# Check stats
memvault stats <user_id>
API Endpoints
| Method | Endpoint | Description |
|---|---|---|
| POST | /memorize |
Store conversation → extract facts/events/knowledge |
| POST | /retrieve |
Strength-weighted vector search (similarity × strength) |
| POST | /decay |
Ebbinghaus forgetting curve (run daily via cron) |
| GET | /stats |
Memory distribution, access patterns, agent breakdown |
| GET | /health |
Service health |
How It Works
- Store: Conversations → LLM extracts facts → embedded → stored in pgvector
- Retrieve: Query embedded → cosine similarity × memory strength → ranked results
- Decay:
strength = exp(-rate × days / (1 + damping × ln(1 + access_count))) - Access boost: Each retrieval increments
access_count, slowing decay
Fading memories (strength < 0.1) are excluded from search.
Configuration
All via environment variables in .env (created by install script):
MEMVAULT_LLM_BASE_URL— Default: Ollama local. Set to OpenAI/Groq/etc URL if preferredMEMVAULT_LLM_MODEL— Default:qwen2.5:3bMEMVAULT_TRANSLATION— Settrue+MEMVAULT_TRANSLATION_LANGfor auto-translationMEMVAULT_PORT— Default:8002
Daily Cron Setup
Add Ebbinghaus decay to your agent's cron:
0 3 * * * curl -s -X POST 'http://127.0.0.1:8002/decay?user_id=YOUR_USER_ID'
TOOLS.md Snippet
## MemVault 🧠
memvault memorize-text "<user_id>" "<content>" "<context>"
memvault retrieve "<user_id>" "<query>"
memvault decay <user_id>
memvault stats <user_id>
- API: 127.0.0.1:8002
Multi-Agent Memory
Tag memories by source agent:
curl -X POST http://localhost:8002/memorize -H "Content-Type: application/json" \
-d '{"conversation": [
{"role": "metadata", "content": "{\"source_agent\": \"research-bot\"}"},
{"role": "user", "content": "Found new papers on transformers"}
], "user_id": "team"}'
Troubleshooting
- "Connection refused" → Run
docker compose -f ~/.openclaw/workspace/skills/memvault/docker-compose.yml up -d - Slow memorize → Normal, LLM extraction takes 5-15s per conversation
- No results from retrieve → Check
memvault stats— if total=0, nothing stored yet - All memories fading → Reduce decay_rate:
curl -X POST 'http://localhost:8002/decay?decay_rate=0.05'
04 同仓库更多 Skills
youtube-content
YouTube transcripts to summaries, threads, blogs.
xurl
X/Twitter via xurl CLI: post, search, DM, media, v2 API.
touchdesigner-mcp
Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals. 36 native tools.
test-driven-development
TDD: enforce RED-GREEN-REFACTOR, tests before code.
teams-meeting-pipeline
Operate the Teams meeting summary pipeline via Hermes CLI — summarize meetings, inspect pipeline status, replay jobs, manage Microsoft Graph subscriptions.
systematic-debugging
4-phase root cause debugging: understand bugs before fixing.