Hey everyone! 👋 I'm Phuc, the maker of Kioku Lite. I've been running AI agents as personal companions and mentors via Telegram. The biggest frustration: they'd forget everything between sessions. I tried existing memory solutions, but they all stored flat text or vectors — no structure, no causal reasoning. So I built kioku-lite: a memory engine that combines BM25 keyword search + vector similarity + a Knowledge Graph, all inside a single SQLite file. No Docker, no cloud APIs, no external databases. The key design choice: kioku-lite never calls an LLM. Your agent (which IS an LLM) extracts entities from its own context and indexes them. This keeps the engine 100% local, LLM-agnostic, and free of per-write API costs. It ships with two built-in personas: 🤗 Companion — tracks emotions & life events 🧭 Mentor — tracks decisions & lessons learned Setup is designed to be dead simple: copy a setup guide and paste it to your agent. The agent installs, configures, and activates kioku-lite autonomously. 📋 Setup guides: → Claude Code / Cursor / Windsurf: https://phuc-nt.github.io/kioku-... → OpenClaw + Telegram: https://phuc-nt.github.io/kioku-... The story behind it: https://phuc-nt.github.io/kioku-... Would love your feedback — especially on the tri-hybrid search approach and how it compares to what you're using for agent memory today!

Kioku Lite

Knowledge Graph memory for AI agents. Single SQLite file.

Knowledge Graph memory for AI agents. Single SQLite file.