Agent Memory — Shared Offline Memory for AI Agents (MCP-native)

[ThePlenkov]

Agent Memory — Docker Desktop Extension

I've built a Docker Desktop extension that gives every AI agent (Claude, Cursor, VS Code Copilot, custom agents) a persistent, searchable memory that works entirely offline — no cloud required.

Repository: https://github.com/ThePlenkov/agent-memory

Why is this useful?

AI coding agents today have no persistent memory between sessions. Every time you start a new conversation, the agent forgets everything — your project conventions, past decisions, debugging context, and lessons learned.

Agent Memory solves this by providing a shared memory layer powered by Hindsight that:

• Runs entirely offline inside Docker Desktop — no cloud, no API keys needed for core functionality
• Works with any MCP-compatible agent — Claude Code, Cursor, VS Code Copilot, or custom agents via standard HTTP MCP protocol
• Persists across sessions — memories survive container restarts via Docker volumes
• Organizes by context — multi-bank support lets you separate memories by project, agent, or topic
• Supports Docker MCP Gateway — compatible with docker mcp gateway connect

What problems does it solve?

1. Agent amnesia: Agents forget project-specific patterns, conventions, and past decisions between sessions
2. Repeated mistakes: Without memory, agents repeat the same errors you've already corrected
3. Knowledge silos: Different agents (Cursor, Claude, Copilot) can't share what they've learned about your codebase
4. Privacy concerns: All memory stays local in Docker — nothing leaves your machine

Architecture

The extension runs as a Docker Desktop extension with:

• React UI — for browsing, configuring, and managing memories
• Go backend — proxies requests to the Hindsight container
• Hindsight container — provides MCP server (port 8888), embedded Postgres, and a web UI (port 9999)

Quick Start

git clone https://github.com/ThePlenkov/agent-memory.git
cd agent-memory
make install

Agents connect via MCP at http://localhost:8888.

Feedback and contributions welcome!