Titen

Open-source agent memory with receipts. No API key, no LLM.

Open Source

Developer Tools

Artificial Intelligence

GitHub

Visit WebsiteSee on Product HuntGithub ⧉

Hey Product Hunt 👋 I'm Rama, maker of Titen. Try this: ask your AI agent "why do you remember that?" Most memory layers can't answer. They hand back a nearby paragraph and a similarity score. That's vibes, not memory. And once a team of agents shares that store, nobody can tell what's current, who was allowed to see it, or which agent quietly overwrote whom. Titen is built on one rule: every memory keeps its receipts. → Where it came from: raw observations with a content hash and a timestamp → Who may read it: org / project / subject scoping runs BEFORE retrieval, not after → What contradicts it: disputes stay disputed, both sides kept, never averaged Three things that make it different: 1. It needs nothing. No API key, no LLM, no embedding provider, zero runtime dependencies. One command (needs Bun 1.2+) and your agent has memory over MCP: npx titen-memory mcp 2. It's a drop-in for @modelcontextprotocol/server-memory. Same 9 tool names, your existing graph is imported on first run. Underneath, search becomes real ranked FTS5 instead of String.includes. 3. The benchmark publishes our losses. On LongMemEval-S (pooled, n=500) the zero-provider lane scored 0.246 recall@1 vs 0.182 for Mem0 OSS (infer=False). Two of our own pre-registered falsifiers also fired against us. They're printed on the chart, not hidden: titen.dev/benchmark 🧾 Launch-day challenge: the Memory Audit Point it at any agent memory store (Titen, the MCP reference server's memory.json, or a Mem0 export): npx titen-memory audit ./memory.json Read-only, offline, nothing uploaded. It counts exact duplicates, near duplicates, recall loops, leaked secret patterns and stale entries, with per-item evidence you can check by hand. We ran it on Titen's own store first: 17.9% duplicates, 96.7% stale. Published in the repo. Your turn. Drop your numbers below 👇 Honest notes: Titen is pre-1.0 (pin an exact version). By default you author the claims; model-assisted extraction ships but stays gated until a model passes 90% on our eval. And on huge unscoped stores, compile latency still crosses our own 250 ms line. Why "Titen"? It comes from the Javanese "niteni": to notice, and to keep what you noticed. The logo is kawung, a four-petal batik motif. Its four petals stand for tenant, subject, agent and run. I'd love to hear: what's the one thing you wish your agent could explain about its own memory? GitHub: github.com/RamaAditya49/titen