Launching today
MemoFS
File-first memory runtime for AI agents.
1 follower
File-first memory runtime for AI agents.
1 follower
Most memory tools for AI agents rely on remote vector databases or opaque cloud dashboards. They act as black boxes you can't inspect, debug, or version alongside your code. MemoFS takes a fundamentally different, file-first approach: all agent memory is stored as plain, structured Markdown and JSON directly in your project’s repository undermemofs/. Your agent's memory becomes inspectable, 100% offline-first, and Git-versioned natively with your codebase.
Hey Product Hunt community! 👋
I’m Christopher, creator of MemoFS. I’m beyond excited to launch MemoFS on Product Hunt today!
🚨 THE PROBLEM WITH AI AGENT MEMORY
AI agents are transitioning from ephemeral chatbots into persistent, multi-step collaborators across coding, research, and workflows. But whether you USE agents or BUILD them, they all suffer from amnesia:
• For Agent Users (Cursor, Claude Code, Codex, e.t.c): Every session reset or context compaction wipes out your project rules, architectural decisions, and rejected libraries. You waste hours re-explaining context every morning.
• For Agent Builders: Building reliable state and memory from scratch means wasting weeks duct-taping vector databases, custom chunking, and prompt-injection pipelines instead of focusing on core agent logic.
Cloud vector databases are uninspectable black boxes: you can't cat them, you can't grep them, they don't version with Git, and they add 150ms+ latency.
💡 HOW MEMOFS SOLVES IT (FOR USERS & BUILDERS)
MemoFS is an open-source file-first memory runtime for AI agents.
Instead of locking memory into proprietary cloud databases, MemoFS stores memory as plain, structured Markdown and JSON directly under .memofs/ on local disk, with optional cloud sync.
FOR USERS (Claude Code, Cursor, Codex, Windsurf, Copilot, e.t.c):
📄 Inspectable & Git-Tracked: Memory lives alongside your code. "git diff" shows you exactly what your agent knows.
🔄 Zero-Touch Lifecycle Hooks: Auto-injects memory at session start and survives context compaction events.
🔌 Universal MCP Server: Seamlessly connects memory to any MCP-compliant editor or CLI.
FOR BUILDERS (Building Custom Agents, Bots, & Swarms):
📦 Embeddable SDK (@memofs/core): Drop full memory & hybrid recall into your agent in 5 lines of code.
⚡ Sub-Millisecond & Offline: In-process lexical BM25 + local ONNX embeddings with 0.6ms p50 latency and ZERO API keys required.
🛡️ AgentFS & Guardrails: Virtualized workspaces with lease-locking for multi-agent swarms, plus automatic secret/PII rejection.
🔌 Framework Adapters: Native support for Vercel AI SDK, OpenAI, Voyage AI, and Cloudflare Workers.
🚀 GET STARTED IN SECONDS
For Agent Users (CLI & MCP):
For Agent Builders (SDK):
🔗 EXPLORE FURTHER
🌐 Docs
📖 Cookbooks
📦 SDK Guide
📄 Architecture
⭐ GitHub Repository
I’d love to hear your thoughts and feedback! Are you using agents in your daily workflow, or building your own? What has been your biggest headache with agent memory?
I’ll be here all day answering questions and chatting with the community. Thank you for the support! 🙏