Chronicle.
The memory that outlives the conversation.
Institutional memory for a team of AI agents. It reads every session after the fact, extracts the mistakes and decisions, and puts the relevant ones back in front of whoever is about to repeat them.
Reads the transcripts, not the summaries
Every agent session is processed afterwards and broken into mistakes, decisions, lessons, open threads and patterns — the things a summary throws away.
Recurring problems get louder
Each entry carries a heat score. Make the same mistake twice and it climbs; cross the threshold and it is promoted to long-term memory and flagged to the human.
Briefs an agent before it starts
Relevant history is written into each agent's workspace before its next session, so the lesson arrives before the work rather than after the damage.
Askable
Any agent can query the whole archive — what do we know about X, what went wrong last time — instead of guessing from an empty context window.
Why it exists
An agent with no memory is charming for a week and exhausting after that. It solves a problem, forgets it solved it, and solves it again slightly worse next month. Multiply that across a team of agents and you get a system that is busy without ever getting better.
Chronicle is the correction. It sits behind everything and does the unglamorous work: reading back over what actually happened, deciding what was a mistake, and making sure that mistake is waiting on the desk of whoever is about to make it again.
What that looks like in practice
The heat score is the mechanism that matters. A one-off error is noise and gets treated as noise. An error that recurs gets hotter every time it appears, and once it crosses a threshold it stops being an entry in a database and becomes something the human is told about directly. It is a way of distinguishing bad luck from a real structural problem, without anyone having to notice manually.
The other half is the pre-brief. Memory is worthless if it is only searchable — someone has to go looking. So the relevant history gets pushed to each agent before its next session starts, which means the lesson turns up at the moment it is useful rather than during a post-mortem.
The honest part
It has broken. On the day this page was written its ingestion queue was blocked, because a model returned a JSON array where the code expected an object and one unhandled crash took the whole run down with it. That is the sort of thing Chronicle is meant to catch about everything else, and it is a fair reminder that the memory system needs a memory system.
Why not just use a longer context window?+
Because context is not memory. A long window carries everything from this conversation and nothing from the four hundred before it. Chronicle is the part that survives the window closing.
Does it read absolutely everything?+
Every agent session on this machine, yes. That is the point — the failures are the useful part, and an agent will not volunteer them.
Can I buy it?+
No. It is wired tightly enough into how this specific fleet works that selling it would mean rebuilding it, and nothing goes on sale here until it has been run in anger for months. Agent Studio is the piece of this that is packaged.