# Keyoku

A cognitive memory engine for AI agents, written in pure Go, that extracts memories from conversations, builds a knowledge graph of entities and relationships, and decays or consolidates memories over time.

Keyoku is a memory engine that sits behind an AI agent, extracting memories from conversations, deduplicating and merging them by meaning, and letting unused memories decay while important ones persist, rather than storing every message as an unchanging row. Every conversation also grows a knowledge graph of people, organizations and how they relate, so the system can connect facts such as one person's employer and manager rather than only recalling that they were mentioned together.

A heartbeat process periodically scans twelve categories of signal across the memory store, such as deadlines, goals and relationship gaps, combines weak signals into decisions and compiles them through an LLM into a single actionable output. It stores data locally in SQLite with an in-process HNSW vector index and has no external dependencies. It is also the storage and memory backend for a separate workflow harness for coding agents, which mirrors its own knowledge into the engine and upgrades its plain-text knowledge queries to semantic search when the engine is running, falling back to local JSONL storage otherwise.

Website: https://keyoku.ai