I built YantrikDB after repeatedly watching capable agents lose the thread between sessions. The first version solved persistence. That turned out to be the easy part. Memory also needs boundaries, history, and a lifecycle. It should be possible to ask why a record surfaced, distinguish an old claim from a current one, keep tenants isolated, and preserve conflicting evidence without silently deleting either side. YantrikDB now ships as one open-source engine you can reach four ways: Rust, Python, MCP, and a network server. The site includes an in-browser WebAssembly lab, a multi-agent example, published retrieval measurements, and the failure report for a conflict detector we deleted after it measured 0/16 precision. The most useful feedback for me is concrete: where does your agent's memory fail today? Missed recall, stale context, namespace leakage, false conflicts, or something else? We also opened a Memory Failure Clinic for small synthetic reproductions, so uncomfortable cases can become public tests instead of vague claims. YantrikDB is Apache-2.0. I would love your questions and hard failure cases.

YantrikDB

Memory that just works for AI agents

Memory that just works for AI agents