Research
Notes on databases, AI agents, and the deterministic-signals bet behind SnoutData.
- What Made Postgres 18 Faster: Larger Reads, Not Async I/OWe measured Postgres 17 against 18 on a host shaped like our fleet, cold cache, a Plus project's limits. One scan got 4.3 times faster, and switching asynchronous I/O off did not slow it. Where the gain comes from, why the sequential scan did not move, and why vacuum got slower.
- Time-Series Storage Inside Postgres, Without a Chunk CatalogSnoutTime keeps Postgres's own partitioning and seals closed partitions into a column store. Measured at 10M, 100M and a billion rows against the Postgres it is built on, with the losses in the same tables as the wins.
- A Rule Is Only a Rule Once It PredictsThe boundary law from the last post was an observation on one backend. We took it to PostgreSQL, a backend it was never fit to, and watched a single error class split along the same line.
- We Built the One Honest OpeningWe built the failure-only, deterministic memory the last post argued for, then measured it: the mechanism, the first numbers, and the flat negative that taught us the most.
- Your Agent's "Memory" Is Not LearningPersistence and learning are different things. Almost everything sold as agent memory is the first one wearing the second one's clothes, and here is the actual line between them.