Vektori
A memory system for AI agents that stores facts, episodes and raw sentences in a three-layer graph backed by a single SQLite database by default, so agents can recall not just what a user said but how a preference or pattern changed over several sessions.
Vektori is a memory library for AI agents built around a three-layer sentence graph: a fact layer for point-in-time statements, an episode layer for the events that produced them, and a raw sentence layer holding the original text a fact or episode traces back to. A search hits the fact layer, the graph then discovers related episodes and traces back to source sentences, so an agent can recover not just a preference but why and how it changed.
It stores data in SQLite by default and can be pointed at Postgres, Neo4j, Qdrant or Milvus when scaling up. On long-horizon memory benchmarks such as LoCoMo and LongMemEval, the project's own measurements put it in a similar range to comparable memory systems such as Mem0, Zep and Letta, and it reports that the right context is already present in the retrieved results for about 95 percent of test questions, with the remaining gap attributed to synthesis rather than retrieval.
It installs as a Python package, with optional extras for Neo4j and other backends.