MemPalace Agent is a small OpenAI-compatible proxy that sits in front of any LLM endpoint and gives it memory.

It does three things:

• loads MemPalace wake-up context
• searches verbatim memory on each request
• writes the conversation back into the palace

Your model endpoint stays yours. MemPalace stays local.

Main module: agent.py

Install

Run

AGENT_LLM_BASE_URL="http://localhost:8000/v1" \
AGENT_LLM_MODEL="your-model" \
uv run agent.py --data ~/my-agent

By default the proxy serves on http://127.0.0.1:8001/v1.

You can also bind it explicitly:

AGENT_BASE_URL="http://0.0.0.0:8889/v1" \
AGENT_LLM_BASE_URL="http://192.168.0.124:8888/v1" \
AGENT_LLM_MODEL="your-model" \
uv run agent.py --data ~/my-agent

Then point any OpenAI-compatible client at:

Example

curl http://127.0.0.1:8001/v1/chat/completions \
  -X POST \
  -H 'Content-Type: application/json' \
  --data-raw '{"messages":[{"role":"user","content":"whats my name?"}],"stream":false}'
  # { "choices": [ { "message": { "role": "assistant", "content": "Your name is Sebastian." } } ] }

Environment

• AGENT_LLM_BASE_URL: backend LLM base URL
• AGENT_LLM_API_KEY: optional bearer token for the backend
• AGENT_LLM_MODEL: model forced by the proxy
• AGENT_LLM_FORCE_MODEL: if true, ignore request model and use AGENT_LLM_MODEL
• AGENT_LLM_EXTRA_HEADERS: JSON object of extra backend headers
• AGENT_BASE_URL: proxy bind URL
• AGENT_HOST: proxy bind host
• AGENT_PORT: proxy bind port

Legacy LLM_BASE_URL, LLM_API_KEY, and LLM_MODEL are also accepted.

Endpoints

• POST /v1/chat/completions
• GET /v1/models
• GET /v1/health
• POST /v1/memory/search
• GET /v1/memory/status

Notes

• The memory store lives under --data/.mempalace/
• Use 127.0.0.1 or a real host when connecting clients; do not send client traffic to 0.0.0.0
• Streaming is supported