Hindsight on Unraid

Shared memory for AI agents, reachable over MCP and HTTP. This template runs the official image ghcr.io/vectorize-io/hindsight:latest, which bundles the API, the web dashboard, the local search models, and an embedded PostgreSQL database. Hindsight also needs one language model, which you point at a local Ollama server or a hosted API.

First start

1. Make the data folders on the Unraid terminal before you start the container:

   mkdir -p /mnt/user/appdata/hindsight
   chown -R 1000:1000 /mnt/user/appdata/hindsight
   chmod 700 /mnt/user/appdata/hindsight
   

   The container runs as user 1000 and does not read Unraid's PUID and PGID settings. If you change a path in the template, use the same path here.

2. Copy templates/hindsight.xml to /boot/config/plugins/dockerMan/templates-user/my-hindsight.xml on Unraid. Then open Docker > Add Container and pick hindsight from the user templates.

3. Set the model fields. The template defaults to local Ollama:

   Field	Value
   Model service	ollama
   Model name	gemma4:12b
   Model endpoint	http://YOUR_UNRAID_IP:11434/v1
   Model provider key	empty

   The model must support tool calling. Pull it on the Ollama host first:

   ollama pull gemma4:12b
   

   For any other service, use these values:

   Model service	Provider	Model	Endpoint	Provider key
   Local Ollama	ollama	Exact local name	http://YOUR_UNRAID_IP:11434/v1	Blank
   OpenAI API key	openai	gpt-4o-mini	Blank	OpenAI key
   Anthropic API key	anthropic	claude-haiku-4-5	Blank	Anthropic key
   Google Gemini API key	gemini	gemini-3.5-flash	Blank	Gemini key
   Groq API key	groq	openai/gpt-oss-120b	Blank	Groq key
   Ollama Cloud key	ollama-cloud	gpt-oss:120b	Blank	Ollama key
   OpenRouter key	openrouter	Exact OpenRouter model id	Blank	OpenRouter key
   DeepSeek key	deepseek	deepseek-v4-flash	Blank	DeepSeek key
   xAI Grok API key	openai	Exact xAI model name	https://api.x.ai/v1	xAI key
   Other OpenAI-compatible API	openai	Exact API name	Provider URL ending in /v1	API key

   Model service also accepts these:

   openai-responses, vertexai, lmstudio, llamacpp, minimax, zai, opencode-go, atlas, meta, volcano, requesty, github-copilot, bedrock, fireworks, nous, litellm, litellmrouter

   The Hindsight models documentation lists the extra variables some of them need.

4. Set the three keys. Memory API key and Web UI backend key must hold the same value. Web UI login key is a separate one.

5. Apply the template. Open http://YOUR_UNRAID_IP:9999/ and sign in with the Web UI login key. Leave the internal Web UI backend URL at http://127.0.0.1:8888.

Connect agents

Upstream ships an installer that detects the coding agents you already have and wires them all up. Run it on the machine you code from, not on Unraid. This has only been run on Linux.

npx @vectorize-io/hindsight-coding-agents install all \
  --server self-hosted \
  --api-url http://YOUR_UNRAID_IP:8888 \
  --api-token YOUR_MEMORY_API_KEY

It covers Claude Code, Codex, Cursor, opencode and Pi, and writes one config file at ~/.hindsight/coding-agent.json. After that each repo gets its own bank, seeded from its commit history.

A repo named myproject lands in a bank named coding-agent::myproject. To send a folder to a bank you named yourself, add a mapping to that same file:

{
  "mapPathToBank": {
    "/path/to/your/repo": "your-bank"
  }
}

Anything the installer does not cover

It is a plain MCP endpoint over Streamable HTTP. One bank per project, named in the URL:

• http://YOUR_UNRAID_IP:8888/mcp/YOUR_BANK/

Send the header Authorization: Bearer YOUR_MEMORY_API_KEY, using the Memory API key from the template. Most tools read the same JSON shape:

{
  "mcpServers": {
    "hindsight-YOUR_BANK": {
      "type": "http",
      "url": "http://YOUR_UNRAID_IP:8888/mcp/YOUR_BANK/",
      "headers": { "Authorization": "Bearer YOUR_MEMORY_API_KEY" }
    }
  }
}

Tool	Where to put it
Cursor	~/.cursor/mcp.json for all projects, or .cursor/mcp.json in one project. Cursor ignores the type field.
Oh My Pi	~/.omp/agent/mcp.json. It expands ${HINDSIGHT_API_KEY} in headers.
Plain Pi	Install the pi-mcp extension, then use ~/.pi/agent/mcp.json with the same shape.
Grok CLI	Run grok mcp add --transport http hindsight-YOUR_BANK http://YOUR_UNRAID_IP:8888/mcp/YOUR_BANK/ --header "Authorization: Bearer YOUR_MEMORY_API_KEY"
Claude Desktop	Its connector dialog has nowhere for a bearer token, so use the mcp-remote bridge in claude_desktop_config.json: "command": "npx", "args": ["mcp-remote", "http://YOUR_UNRAID_IP:8888/mcp/YOUR_BANK/", "--header", "Authorization: Bearer YOUR_MEMORY_API_KEY"]
ChatGPT Desktop	Needs a public HTTPS server with OAuth or no auth, so a LAN address will not work.

The shared API key opens every bank, so only give it to clients you trust.

Storage and updates

The appdata path maps to /home/hindsight/.pg0 inside the container. Keep that mapping on updates. Before an update, stop the container and back up the whole mapped folder.

To move to an external PostgreSQL with pgvector, export and import your memories first, then change Database URL under Show more settings from pg0 to the connection string. Changing the URL alone does not move existing memory. The external database must already exist with its vector extension enabled before Hindsight starts.

Sources

• Docker setup and storage
• Model configuration and supported services
• API and dashboard keys
• MCP connections
• Releases