OpenClaw Advanced Memory

Provides persistent, searchable AI agent memory with real-time capture, vector search, and nightly LLM curation for long-term recall on local hardware.

Three-tier AI agent memory system — real-time capture, vector search, and LLM-curated long-term recall.

What It Does

Gives your OpenClaw agent persistent, searchable memory that survives across sessions:

• HOT tier — Redis buffer captures conversation turns in real-time (every 30s)
• WARM tier — Qdrant vector store with chunked, embedded conversations (searchable, 7-day retention)
• COLD tier — LLM-curated "gems" extracted nightly (decisions, lessons, milestones — stored forever)

Requirements

• Qdrant — vector database (Docker recommended)
• Redis — buffer queue (Docker recommended)
• Ollama — local embeddings (snowflake-arctic-embed2) + curation LLM (qwen2.5:7b)
• Python 3.10+ with qdrant-client, redis, requests

No cloud APIs. No subscriptions. Runs entirely on your own hardware.

Setup

# 1. Start Qdrant + Redis (Docker)
docker compose up -d

# 2. Pull Ollama models
ollama pull snowflake-arctic-embed2
ollama pull qwen2.5:7b

# 3. Run the installer
bash scripts/install.sh

The installer sets up Qdrant collections, installs a systemd capture service, and configures cron jobs.

Edit connection hosts at the top of each script if your infra isn't on localhost.

Usage

# Search your memory
./recall "what did we decide about pricing"
./recall "deployment" --project myproject --tier cold -v

# Check system status
./mem-status

# Force a warm flush or curation run
./warm-now
./curate-now 2026-03-01

Schedules

Component	Schedule	What It Does
mem-capture	Always running (systemd)	Watches transcripts → Redis
mem-warm	Every 30 min (cron)	Redis → Qdrant warm
mem-curate	Nightly 2 AM (cron)	Warm → LLM curation → Qdrant cold

How Curation Works

Every night, a local LLM (qwen2.5:7b via Ollama) reads the day's conversations and extracts structured gems:

{
  "gem": "Chose DistilBERT over TinyBERT — 99.69% F1, zero false positives",
  "context": "A/B tested both architectures on red team suite",
  "categories": ["decision", "technical"],
  "project": "guardian",
  "importance": "high"
}

Only decisions, milestones, lessons, and people info make the cut. Casual banter and debugging noise get filtered out.

Links

• GitHub: https://github.com/jtil4201/openclaw-advanced-memory
• Full docs: See README.md in the repo for architecture diagrams, tuning guide, and adaptation notes