I built Smrti because I kept hitting the same wall: agents forget everything between sessions. The usual fix is a vector DB, but similarity search alone doesn't tell you how confident a memory is, how it connects to other things the agent knows, or whether it's something the agent should actively avoid repeating. So Smrti stores memories as graph nodes — concepts, beliefs, episodes, goals — each carrying a Bayesian truth value, emotional valence, and attention weight. The whole thing is one SQLite file. sqlite-vec handles KNN, embeddings run on CPU via ONNX. Nothing else to deploy. Retrieval goes beyond cosine similarity. After the vector search it walks one hop through the graph, then scores everything by a salience formula that mixes similarity, attention, confidence, and emotional intensity. The bit I'm most pleased with: memories of past screwups (outages, antipatterns) get dynamically boosted so they surface even when they're old. Results come back tagged critical_warning, known_antipattern, or context. There's a background consolidation loop that merges new evidence using PLN, decays attention, promotes important stuff to long-term storage, and weakens the less confident side when two memories contradict each other. A personality system with 16 hyperparameters controls all the weights. I ship six presets (balanced, analytical, curious, empathetic, maverick, deterministic) but you can tune individual parameters if the presets don't fit. Three serving modes: MCP for Claude and similar, a REST API, and an OpenAI-compatible proxy you can drop in front of any chat completions endpoint. The proxy injects relevant memories into the system prompt automatically. Honest caveat: the entity resolution (exact match, alias, fuzzy, embedding fallback) works well enough for my use cases but I haven't stress-tested it on anything with tens of thousands of entities. If you try it at that scale I'd like to hear how it goes.

smrti

Memory engine for AI agents

Memory engine for AI agents