Kimi Code Memory

A Python MCP (Model Context Protocol) bridge that exposes TencentDB Agent Memory as MCP tools for Kimi Code CLI.

What it does

Gives your AI coding assistant long-term memory across sessions:

• L0 - Raw conversation storage
• L1 - Atomic memory facts (auto-extracted)
• L2 - Scene/context blocks (auto-clustered)
• L3 - User persona/profile (auto-generated)

The LLM can recall relevant memories, capture new conversations, and search past interactions.

Architecture

Kimi Code CLI  <---stdio--->  Python MCP Bridge (this repo)
                                   |
                                   | HTTP :8420
                                   v
                          TencentDB Agent Memory Gateway
                          (official npm package, runs locally)

This repo is a thin Python bridge — it forwards 5 MCP tools to the official Gateway via HTTP. The Gateway does all the heavy lifting (L0-L3 extraction, vector search, persona generation).

Quick Start

1. Install Python dependencies

pip install -r requirements.txt

2. Set up the Gateway (one-time)

python setup-gateway.py

This installs the official @tencentdb-agent-memory/memory-tencentdb npm package and tsx into ~/.memory-tencentdb/.

3. Configure credentials

cp .env.example .env
# Edit .env and fill in your API keys

You need:

• LLM API key — any OpenAI-compatible endpoint (SiliconFlow, OpenAI, SenseNova, etc.)
• SiliconFlow API key — for embeddings (BAAI/bge-m3)

4. Start the Gateway

python start-gateway.py

For background/autostart mode:

python start-gateway-background.py

5. Register in Kimi Code

Add to your ~/.kimi-code/mcp.json:

{
  "mcpServers": {
    "tencentdb-memory": {
      "command": "python",
      "args": ["path/to/server.py"]
    }
  }
}

6. (Optional) Auto-invoke on every conversation with AGENTS.md

If you want Kimi Code to automatically recall memories at the start of every conversation and capture them after each turn, create an AGENTS.md file in your project root. AGENTS.md is a project-scope instruction file that Kimi Code loads automatically; it is not a skill and does not require a trigger word.

Example AGENTS.md:

# TencentDB Agent Memory Rules

- **Fixed session_key:** always use `kimi-default` (or any stable identifier).
- **On the first user message of every conversation:**
  - Call `mcp__tencentdb-memory__tencentdb_memory_recall` with `session_key="kimi-default"`.
  - Do not answer the user until this recall has been attempted.
- **After every meaningful user/assistant turn:**
  - Call `mcp__tencentdb-memory__tencentdb_memory_capture` with `user_content`, `assistant_content`, and `session_key="kimi-default"`.
- **When the session ends:**
  - Call `mcp__tencentdb-memory__tencentdb_session_end` with `session_key="kimi-default"`.

This is useful for keeping a single, persistent memory context across all your chats in a workspace.

MCP Tools

Tool	Description
tencentdb_memory_recall	Recall relevant L1/L2/L3 memories for current query
tencentdb_memory_capture	Store a completed conversation turn into memory pipeline
tencentdb_memory_search	Search structured memories (L1-L3) with optional type filter
tencentdb_conversation_search	Search raw L0 conversation history
tencentdb_session_end	Flush pending extraction work for a session

SKILL.md

Include SKILL.md in your Kimi Code skills directory to teach the LLM when and how to use these memory tools.

Requirements

• Python >= 3.12
• Node.js >= 22.16.0 (for the Gateway)
• An OpenAI-compatible LLM API key
• A SiliconFlow API key (for embeddings)

Acknowledgments

• TencentDB-Agent-Memory by TencentCloud (MIT License)
• FastMCP Python framework

License

MIT — see LICENSE

This project includes modifications based on TencentDB-Agent-Memory by TencentCloud.