chromadb-memory（色度数据库内存）

ChromaDB Memory

Long-term semantic memory backed by ChromaDB and local Ollama embeddings. Zero cloud dependencies.

What It Does

• Auto-recall: Before every agent turn, queries ChromaDB with the user's message and injects relevant context automatically
• chromadb_search tool: Manual semantic search over your ChromaDB collection
• 100% local: Ollama (nomic-embed-text) for embeddings, ChromaDB for vector storage

Prerequisites

1. ChromaDB running (Docker recommended):

   docker run -d --name chromadb -p 8100:8000 chromadb/chroma:latest

1. Ollama with an embedding model:

   ollama pull nomic-embed-text

1. Indexed documents in ChromaDB. Use any ChromaDB-compatible indexer to populate your collection.

Install

# 1. Copy the plugin extension
mkdir -p ~/.openclaw/extensions/chromadb-memory
cp {baseDir}/scripts/index.ts ~/.openclaw/extensions/chromadb-memory/
cp {baseDir}/scripts/openclaw.plugin.json ~/.openclaw/extensions/chromadb-memory/

# 2. Add to your OpenClaw config (~/.openclaw/openclaw.json):

{
  "plugins": {
    "entries": {
      "chromadb-memory": {
        "enabled": true,
        "config": {
          "chromaUrl": "http://localhost:8100",
          "collectionName": "longterm_memory",
          "ollamaUrl": "http://localhost:11434",
          "embeddingModel": "nomic-embed-text",
          "autoRecall": true,
          "autoRecallResults": 3,
          "minScore": 0.5
        }
      }
    }
  }
}

# 4. Restart the gateway
openclaw gateway restart

Config Options

How It Works

1. You send a message
2. Plugin embeds your message via Ollama (nomic-embed-text, 768 dimensions)
3. Queries ChromaDB for nearest neighbors
4. Results above minScore are injected into the agent's context as <chromadb-memories>
5. Agent responds with relevant long-term context available

Token Cost

Auto-recall adds ~275 tokens per turn worst case (3 results × ~300 chars + wrapper). Against a 200K+ context window, this is negligible.

Tuning

• Too noisy? Raise minScore to 0.6 or 0.7
• Missing context? Lower minScore to 0.4, increase autoRecallResults to 5
• Want manual only? Set autoRecall: false, use chromadb_search tool

Architecture

User Message → Ollama (embed) → ChromaDB (query) → Context Injection
                                                  ↓
                                          Agent Response

No OpenAI. No cloud. Your memories stay on your hardware.