Agent memory under a million rows: a timestamp column beats top-k

A hundred thousand rows of 384-dimension float32 embeddings, the size the post's own example generates ^[8], is 153.6 MB of vectors that a brute-force scan has to read on every query ^[13]. The quoted 5 to 20 milliseconds per query ^[7] therefore implies read bandwidth somewhere between 7.7 and 31 GB/s ^[14]. The slow end of that is an unremarkable single-threaded numpy pass over an in-memory buffer. The fast end is not, unless the table sits well below the stated ceiling or the scan is spread across cores. The post ships working code but no hardware, dataset or timing method behind the number ^[18], so treat it as an order of magnitude rather than a result: a few thousand memories are free, a hundred thousand cost real milliseconds. Even the slow end stays inside a network round trip, which is the author's basis for saying networked vector stores rarely win on individual queries [c7b].

Latency is not the strongest part of the case. The WHERE clause is. A predicate on tag and timestamp returns every row that satisfies both and nothing else, one row out of three in the worked example ^[4], with no approximation step and no hybrid postfilter to reason about ^[16]. The equivalent against a vector store is top-k with filtering bolted on before or after the ANN lookup, or a query-and-filter loop you maintain yourself ^[6], and what comes back depends on k and on how the index handles the predicate. According to the author, that recall also moves when embeddings drift or the embedding model is upgraded ^[5]. A timestamp comparison does not move. When an agent misbehaves, the gap between those two properties is the gap between reconstructing the context it had and estimating it.

What the SQL route actually charges you is a schema decision taken at insert time. The author's model of agent memory has a recent cache, a working or episodic slice, and a long-term archive rather than one flat event log ^[17], which in the proposed table becomes a `level` column on `memory_event` alongside `agent_id`, `tag` and `embedding` ^[11]. Something has to classify each memory as it is written, and the compound index on `(agent_id, level, timestamp)` ^[11] only earns its keep if queries name all three. Determinism has a price and that is where it is paid: the shape of recall is fixed in advance, in code you can read, instead of by tuning k until the answers look reasonable. The author is explicit that this trade only applies to layered episodic memory, and that pure semantic retrieval is what vector stores are for ^[10].