Vektori

Vektori addresses a real pain point in AI agents: memory that goes beyond simple fact retrieval. Its three-layer graph (Facts, Episodes, Sentences) lets agents recall not just what a user said, but how preferences evolved over time. This is a genuine improvement over flat vector stores where context is lost. We'd reach for this when building a support bot that needs to track changing user preferences, or a multi-agent system needing shared memory. The sentence-level splitting preserves semantic boundaries, and the three retrieval depths (L0–L2) give you control over context window costs. The dual storage (vector + graph) is clever—vector search seeds candidates, graph traversal enriches them with relationships. Where it bites: Vektori is purely self-hosted. There is no managed cloud tier, no WebUI, and no plug-and-play adapter for popular agent frameworks like LangChain or CrewAI (though you can wire it yourself). Benchmark scores vary: LoCoMo 66% and LongMemEval-S 73% are decent but not state-of-the-art; the 95% hit rate is for 'answer in retrieved context,' not exact answer rate. The API is Python-first with no JavaScript SDK yet. Compared to Mem0, which offers a cloud-hosted option and broader framework integrations, Vektori is more flexible and transparent (Apache 2.0), but requires more effort to operate. Compared to LangMem (LangChain's memory module), Vektori provides structured graph traversal rather than simple conversation buffering. In practice, if you need a managed solution or have a non-Python stack, look elsewhere. If you're comfortable running your own infrastructure and want memory with relational awareness, Vektori is worth the setup.