Giving AI Agents Real Memory with Cloudflare Vectorize

Your AI agent forgets everything between sessions. Here's how we fixed that for $6/month.

The Memory Problem

AI assistants are goldfish. Ask Claude or GPT something today, and tomorrow it's gone. OpenClaw and similar agent frameworks inherit this — sessions are isolated, context doesn't persist.

This means:

• Repeating preferences every conversation
• Re-explaining project context
• Agents making the same mistakes twice
• No institutional knowledge building over time

The Solution: Semantic Vector Memory

We built a memory layer using:

• Cloudflare Vectorize — Vector database for embeddings
• Workers AI — bge-base-en-v1.5 for embedding generation
• R2 — Persistent storage for raw memory files
• Workers — REST API for query/ingest

Total cost: ~$6/month for multiple agents with thousands of memories.

How Auto-Capture Works

After every agent response, a hook analyzes the conversation for:

const patterns = {
  decision: /decided|chose|going with|settled on/i,
  correction: /actually|no,|wrong|correction/i,
  preference: /prefer|always|never|like|hate/i,
  learning: /learned|TIL|realized|discovered/i
};

Matches get embedded and stored automatically. No manual "remember this" commands.

How Auto-Recall Works

Before every response:

1. User message gets embedded
2. Vectorize returns top-k similar memories
3. Memories inject into system context
4. Agent responds with historical awareness

Query: "What did we decide about storage?"
→ Vector search (cosine similarity)
→ Returns: "Decision: Triple-layer storage - Git + D1 + R2"
→ Agent responds with context it "remembers"

The Architecture

┌─────────────────────────────────────────────┐
│                 OpenClaw                     │
├─────────────────────────────────────────────┤
│  ┌─────────┐    ┌──────────┐    ┌────────┐ │
│  │ Session │───▶│ Memory   │───▶│ Vector │ │
│  │         │◀───│ Plugin   │◀───│ Worker │ │
│  └─────────┘    └──────────┘    └────────┘ │
└─────────────────────────────────────────────┘
                       │
         ┌─────────────┼─────────────┐
         ▼             ▼             ▼
    ┌─────────┐  ┌──────────┐  ┌─────────┐
    │   R2    │  │Vectorize │  │Workers  │
    │ (files) │  │ (index)  │  │   AI    │
    └─────────┘  └──────────┘  └─────────┘

Results

• 39 indexed memories across 2 agents
• 200-400ms query latency
• 0.95 threshold prevents duplicate storage
• Cross-agent search — agents can access each other's memories
• Type filtering — query only decisions, or only corrections

Why This Matters

AI agents that remember become genuinely useful assistants. They:

• Build institutional knowledge
• Learn from corrections
• Maintain consistent preferences
• Avoid repeating mistakes

This is table stakes for production AI systems. The fact that most frameworks don't include it out of the box is wild.

Built with OpenClaw + Cloudflare. The memory layer that makes AI actually useful.