Vector-store memory as the agent's long-term memory

The original design was the obvious one: embed every turn of every session, store it in pgvector, retrieve top-k at each step. It was the default in three delivery patterns from February to June and it worked for agents with a handful of sessions. The field opened on the conviction that this shape would not survive contact with a long-running workload, and the bench was proposed to find out where it broke.

It broke at roughly 50k tokens of accumulated history. Retrieval latency crossed generation latency, and recall precision on the decision-consistency split fell to 0.61 because the top-k chunks were the ones that looked like the question rather than the ones that held the decision. Adding a re-ranker bought eight points and doubled the latency. The four-way bench ran for eight weeks and the episodic-plus-summary design won on both axes; the write-up became r-memory-layer.

This is a superseded entry, not a refuted one. Vector recall is still the right shape for short-lived agents and for semantic lookup inside the semantic store. What died is the claim that it is the long-term memory. Anyone reaching for it as a default for a multi-session agent should read the recommendation first.