Until now the system started every session as a blank slate — picture an employee who's brilliant but shows up each morning having forgotten everything about the job. Day 3 fixed that with real memory, in two parts. First, a place to store durable facts so they survive a restart. Second, a 'router' that decides, for any given question, where the answer should come from: long-term facts, the current live state of things, or a log of past events. The router matters because not every question wants the same kind of memory — 'what's our standard setup' is a stored fact, 'what's running right now' is live state, and 'what did we do last week' is history. Sending each question to the right place is both faster and more accurate than dumping everything into one pile and hoping.
Done
- Set up storage that keeps facts alive across restarts, plus a layer that reads raw information and turns it into clean, individual facts it can recall later. Without this, the system re-learns the same things every session — which is slow and means it can make the same mistake twice.
- Loaded in the system's core facts about its own setup and confirmed it could recall them accurately when asked in different ways.
- Built the three-way router — durable facts, versus live state, versus past events — that picks the right source using simple rules instead of asking an AI model every time. That keeps it fast and essentially free to run.
- Confirmed the system now pulls the right facts on its own when it needs them, instead of us having to paste those facts into its instructions by hand.
Issues
- Hit a subtle misconfiguration where the memory layer was quietly sending its work to the wrong place behind the scenes — it looked like it was working but wasn't actually storing anything. Found it and fixed it. We also replaced the part doing the fact extraction after the first choice kept producing garbled output and stalling; the replacement did the same job in a fraction of the time.
Deferred
- Plugging the memory router straight into the agent's live runtime — It's a safe, purely additive change — but the running agent is stable, and you don't modify a stable system at the end of a session just because you can. It gets done as its own clean, deliberate step.
Tomorrow
- Add a safety net so the system can record what it's doing and recover cleanly if it ever crashes.
Worth knowing
- Memory is the line between a tool you have to re-explain every single time and a system that already knows the state of the shop. It's the biggest single jump toward something that genuinely runs on its own.
Quota
Compute budget: nominal.
← all days