Every AI session starts from zero. You re-explain context. The agent re-asks. It suggests what you already tried. Agentmemory.md is a self-hosted MCP memory server that persists across every session. Works with Cursor, Claude Desktop, OpenClaw, ChatGPT and any MCP agent. What it stores: decisions with rationale, project context, failed experiments, goals, people, coding preferences. Docker + Tailscale. One-click install for Cursor.

MongoDB Atlas — Build AI apps, not pipelines. Unified data on Atlas.

Build AI apps, not pipelines. Unified data on Atlas.

I built this for myself after months of re-explaining the same context to Cursor every single morning. The mental model I landed on: a Zettelkasten you build together with your agent. Not a note-taking app bolted onto an LLM — a shared brain where both of you write in and both of you benefit from. Under the hood: knowledge graph (PostgreSQL + Apache AGE), hybrid search (BM25 + vector + graph boost), 22 MCP tools, local embeddings — your data never leaves your machine. The retrieval is the interesting part. Not every memory is equal — nodes are scored by similarity × 0.80 + graph proximity × 0.15 + recency × 0.05. Decisions surface before preferences. Recent context beats old noise. Graph edges connect people to projects to decisions so you get lateral recall, not just keyword matches. Happy to answer questions about the architecture, the MCP design, or self-hosting setup.

MongoDB Atlas — Build AI apps, not pipelines. Unified data on Atlas.

Build AI apps, not pipelines. Unified data on Atlas.

Persistent memory for AI agents

Works with Cursor, Claude, OpenClaw, any MCP-comp agent

Works with Cursor, Claude, OpenClaw, any MCP-comp agent