Every event carries a short summary — the line the next agent reads — the structured detail behind it, the files it touched, and a salience score from 0 to 5.

Salience is the product's central opinion. A decision made an hour ago outranks a file read from a minute ago, so file_read is 0 while decision, blocker and usage_limit are 5. That is what makes a small brief useful rather than merely short.

The log is append-only in the database, not just in the application. No interface can rewrite what an agent did.

No model call. The same log and the same budget produce the same brief, byte for byte — a brief you cannot reproduce is one you cannot debug, and an agent that gets a different story on each reconnect is worse off than one that gets none.

The budget is spent from most to least valuable, so what survives a tight one is what a fresh agent would most regret not knowing.

A conclusion routes the next agent. Only the conversation explains why three approaches were abandoned, and that is the expensive thing to rediscover.

Reasoning is stored whole, because it exists nowhere else once the process exits. Tool output is cut to its head and tail, because a file's contents are one tool call away. A failing result is promoted — it is the evidence for a dead end.

Search runs over keywords and meaning together: exact identifiers are what agents actually look for and what embeddings are worst at, and the paraphrase is what keywords miss.

Agents check in as they work. One call states what they are doing, returns what changed since last time, renews their claims and reports conflicts.

Claims are advisory and named accordingly — nothing can intercept another agent's edit tool, so the value is entirely in the other agent being told. They expire on their own, because an agent killed by a usage cap never releases anything.