I built "Lucy" — a personal AI agent that actually remembers.
As a technical artist and game dev, I got curious: how useful does an AI agent get if it never forgets? So I self-hosted one on the Claude Agent SDK, running 24/7 and reachable over Telegram, and spent most of the effort on the hardest part: memory.
Lucy's memory is designed in five tiers, inspired by research from Mem0, Zep and Letta:
— Live today: hybrid recall (full-text + 768-dim vectors) that auto-injects relevant memory, plus episodic memory that stores individual conversation turns
— Built and kept behind a dry-run gate: Mem0-style self-consolidation and Zep-style bi-temporal forgetting
Nothing rewrites memory without my sign-off, and all of it runs on a 1.9 GB VPS.
Beyond memory, Lucy has a "hand": a tool registry + MCP for files, web and git, a safety rail that blocks dangerous commands, and an engine that takes a task and runs it to a report on its own.
The biggest lesson: what turns a chatbot into a colleague isn't the model, it's the memory, the tools, and the safety rail around it.






