Mnemonic is my open-source layered memory system for AI agents. This page is a stripped-down sandbox that runs entirely client-side so you can play with the ideas without cloning the repo. Store memories, watch them get auto-classified into layers, run semantic queries, see contradictions get flagged. Everything you type stays in your browser.
Short-term working memory. Things you just said or noticed. Recent, decayable, often ephemeral. Triggers on temporal markers ("just", "right now", "today").
Specific events with time, place, or people. "Met Sarah yesterday" or "Shipped v2 last week." Indexed by when and where, not by meaning.
Stable facts and preferences. "I prefer dark mode" or "Python uses whitespace for scoping." These don't decay; they're the base layer of what-is-true.
Named entities and how they connect. "Sarah works at Acme Corp." The relational layer is what lets an agent reason about who and what, not just what happened.
In production Mnemonic, classification uses an LLM call with a calibrated prompt. Here we use keyword rules for speed. Querying uses real embedding similarity in production; here we use TF-IDF. The architecture is the same; the implementation is simpler. See the full source →