Hey Product Hunt 👋 ▎ ▎ I built verifiable-memory out of frustration: every agent I worked with eventually ▎ did the same thing — answered a question confidently and wrong, or "remembered" ▎ something it was never told. For anything serious (legal, finance, health, ▎ automation) that's not a bug, it's a liability. ▎ ▎ So I went the opposite way. Instead of making the model guess better, I gave it a ▎ memory that can only answer from facts it actually has — with the source cited — or ▎ honestly says "I don't know." No fabrication by construction, not by prompt. It ▎ also does what an LLM can't from its weights: provable forgetting (GDPR), ▎ valid-time ("what did we know as of date T?"), Merkle proofs, and fully ▎ deterministic recall. ▎ ▎ I stress-tested it to 1,000,000 facts on a plain CPU box: 0% hallucination on ▎ adversarial traps, ~4µs/query, no GPU. It ships as an MCP server + Python SDK, so ▎ it drops in behind Claude or any agent in a couple of lines. ▎ ▎ It's open-source (MIT) and free to self-host: pip install verifiable-memory-mcp · ▎ github.com/Mars-proj/verifiable-memory ▎ ▎ Honest scope: it's a trust/memory layer, not a smarter chatbot — it wins on ▎ verifiability, not fluency. Pair it with your LLM. ▎ ▎ I'd love feedback, especially from anyone building agents where a made-up fact ▎ actually costs something. What would make this useful for you? ▎ ▎ — Sergey

verifiable-memory

Memory for AI agents that can't hallucinate

Memory for AI agents that can't hallucinate