A model plugs in: hooks snap on, tools attach, an agent record appears, and the boot handshake pulls the last savepoint, recent messages, the role contract, and what matters from memory.
The store: a cloud of projected points coloured by kind. A write lands — a row arrives as text, the embedder folds it into one point, and it drops where it belongs, while bulk rows wait in a backlog and a query jumps the line.
Recall, every turn: a message becomes a point, a similarity radius widens while the keyword lane lights literal matches, the sets merge, and the top five rise out of the cloud and materialise as text in the prompt. Then the model answers.
Write, every turn: the answer lands on the floor, facts are extracted and reconciled — add, supersede, or keep — a gate's verdict becomes a lesson row with its provenance tags, and a savepoint closes the session with fork notes.
Check on the system: an agent reads the system in turn, sends one message with a heartbeat, and a second agent's first recall finds what the first one wrote. Closing caption: the network is the memory.