MemoFS

What it solves

AI coding agents currently keep memory in ad hoc, single-purpose files, a CLAUDE.md here, a set of Cursor rules there with no shared format, no versioning, and no way to carry context between tools. Every new session, or every switch from Claude Code to Cursor to Codex, means the agent starts cold: project conventions, prior decisions, and past context have to be re-explained by hand, or are silently lost if no one thought to write them down anywhere durable. Teams building their own coding agents hit the same problem from the infrastructure side — they end up hand-rolling storage, chunking, embeddings, and retrieval just to give their agent any persistent memory at all, usually landing on a proprietary, uninspectable store bolted onto a vector database. MemoFS solves both. For people using coding agents day to day, it gives every agent a shared, persistent memory that survives session resets and tool switches. For people building agents, it's a ready-made, provider-neutral memory runtime — file-based, versioned, and diffable — so they integrate memory instead of building it from scratch.

MemoFS is a file-first memory runtime for AI agents — the same layer whether you're an engineer running Claude Code, Cursor, or Codex day to day, or a team building your own agent and need memory infrastructure to embed rather than write from scratch. At the center is @memofs/core, a provider-neutral engine that stores memory as structured files under .memofs/ instead of a database or vector store: core.md holds the canonical facts loaded every session, notes.md holds longer-form context pulled in on demand, JSONL logs record every memory write and conversation, a recall index of chunks + embeddings powers hybrid search, a versioned entity graph tracks relationships, and snapshots let you roll back a bad memory write. Eleven canonical files, all plain text, all diffable, all sitting in your repo instead of a store you can't inspect. Retrieval blends BM25, fuzzy matching, and semantic search (OpenAI, Voyage AI, or local ONNX embeddings via Transformers.js), plus entity-graph extraction — so an agent gets relevant context back, not just keyword hits. Distribution is dual-path: lifecycle hooks for agents that support them (Claude Code, Codex, opencode), and MCP for the ones that don't (Cursor, Gemini, Copilot) — one integration, either way. The core runtime runs on Node.js 22+ or Cloudflare Workers, is fully open source under MIT, and deliberately never bundles provider-specific code, cloud tenancy, or pricing logic — that's kept out at the adapter and Cloud layer. MemoFS Cloud adds hosted sync, a managed MCP endpoint, and team workspaces on top: free for individuals, paid tiers for teams and enterprises.