Giving Claude Code a memory that outlasts the terminal

What it does

claude-mem is a persistent memory layer for Claude Code, Gemini CLI, OpenCode, and similar agents. It hooks into session lifecycle events to record tool usage and observations, stashes them in a local SQLite database with Chroma vector search, and automatically primes new sessions with relevant history. The goal is continuity: your agent remembers yesterday’s decisions without you repeating them.

The interesting bit

Instead of dumping entire chat logs into the context window, the project treats retrieval as a token-budget problem. It exposes a three-layer MCP search workflow—index, timeline, then full observation—that the documentation claims yields ~10× token savings by filtering before fetching details. There is also a local web viewer on port 37777 if you want to browse your agent’s memory like a security DVR.

Key highlights

• Hooks into session lifecycle events (SessionStart, PostToolUse, SessionEnd, and others) to record observations automatically
• Stores data locally in SQLite with hybrid semantic and keyword search via Chroma
• Surfaces memory through a local web viewer and MCP tools with progressive disclosure layers
• Supports <private> tags to keep sensitive content out of the persistent store
• Also integrates with OpenClaw gateways and Claude Desktop

Caveats

• Installing via npm install -g only gets the SDK, not the plugin hooks or worker service; the npx route is required for full functionality
• Experimental features like Endless Mode live in a beta channel
• The architecture depends on Bun and Node ≥ 18

Verdict

Worth a look if you live in Claude Code or Gemini CLI and are tired of re-establishing context each morning. Skip it if you prefer stateless agent interactions or don’t want another local service running on port 37777.