KektorDB
A memory system for AI agents that pairs vector search with a temporal knowledge graph and a background process that consolidates memories, flags contradictions and lets unused facts fade over time.
KektorDB combines high-performance vector search using HNSW with a temporal knowledge graph and what it calls a cognitive engine, which continuously analyzes stored data, detects contradictions, tracks importance, and lets irrelevant information fade naturally rather than growing without bound.
A background process it calls the Gardener consolidates duplicate memories, detects when a new fact contradicts an established one, identifies gaps in the knowledge graph, and can propose or, in its more autonomous mode, automatically resolve minor contradictions. A separate scoring system assigns each memory a confidence level of crystallized, stable, volatile or contested, based on how consistently it has been confirmed. Once set up, an agent connected over MCP gains dozens of tools covering memory storage and recall, graph traversal, session management and confidence queries.
It is set up and run through a single command-line tool.