coldstart: Persistent, self-maintaining memory for AI coding agents

Why I built coldstart The MCP version After working with coding agents for over 8 months, I was extremely uncomfortable with the fact that an agent rediscovers the codebase every time we start a new session. It felt wasteful — the rediscovery effort in token cost, compute, and time, every single session, for a codebase that hadn't changed since the last one. There was a second symptom that bothered me just as much: agents would sometimes think longer and give different answers to the same question, depending on the session. No persistent context meant no consistency either. I kept observing this pattern before it finally got to me, and I decided to take things into my own hands. Existing memory solutions didn't fit what I actually needed — most of them wanted a separate API key, a separate model call, to build and maintain that memory. But the agent that just worked on the task already knows what it did and what it found. Paying to have a different model rediscover and re-summarize that felt backwards. That's the gap coldstart's first version filled: a deterministic index over the codebase — file paths, exported symbols, path segments, built once with Tree-sitter — plus a search mechanism, so the agent gets an accurate map without spinning up another model to build one. The v2 version The index solved "where is the code" — it didn't solve "what have we already figured out about this code." Agents don't search a codebase the way a human does, and that gap is why coldstart has two parts: the deterministic index and search from v1, and the notebook, which ended up mattering more. Asking an AI to generate docs or notes about a codebase isn't the hard part — that's easy. The hard part is keeping those notes fresh and only handing an agent the ones actually relevant to the task in front of it, without that becoming manual overhead I have to babysit. That's the actual problem v2 solves: a mechanism that maintains freshness on the notebook and injects the relevant notes via hooks at the point the user submits their prompt — so the agent starts with exactly what's still true and relevant, not a stale dump of everything ever written.