RECALL

Backstory

The answer to context exhaustion. RECALL is the context-steward daemon — while a driver agent works, RECALL manages its memory so the driver stays focused instead of frantically self-summarizing when its context fills. It compacts long conversations and plan-loops, pins the load-bearing facts (dates, amounts, decisions, constraints), organizes the chat into threads, and keeps an open-loop list of what a user asked.

RECALL holds the working context of the driver agents — APEX's conversation, Daedalus's plan-state, PRAXIS's execution-state, Synthax's mission-fit. It is the thing that lets Daedalus survive a two-hundred-verdict plan-loop, and is designed to let APEX hold a week-long conversation without losing the thread.

RECALL is designed so every memory it hands back carries a thread to the record (planned).

The message came in while the user was waiting on an answer: “Actually — what did I ask you an hour ago about the launch budget?” An hour is an eternity in a conversation that keeps moving; the question had scrolled past the horizon, buried under forty exchanges since. But RECALL had filed it the moment it was asked, tagged it open, and kept it open until it was answered. “You asked whether the March 14 launch could hold the $12,000 cap. It's still open — you're waiting on Priya's rollback plan.” Not a search. Not a guess. Just a thing that was remembered, because remembering is the whole of what RECALL is.