Agent memory as one SQLite file: 30µs recall, model-free self-improvement loop

inbask · reddit · 2026-09-03

The author open-sources Areev (Rust, 17 crates, dual MIT/Apache): agent memory as a plain SQLite file on disk, recall as a function call, and a self-improvement loop that never calls a model.

Structural recall: 30µs p50 in-process; the same recall via a localhost HTTP sidecar costs 158µs — the network hop is the cost, not the store;
Runs on old hardware: a $35 Raspberry Pi 3 (2016) serves recall at 361µs, flat from 500 to 8,000 grains; a 2018 Intel NUC matches a 2024 laptop at 30µs — all measured, not extrapolated;
Model-free improvement loop: 13 deterministic analyzers read the agent's execution history and emit typed recommendations citing evidence by content hash; human approval, invertible applies, re-measured at 1d/7d/30d with auto-proposed reverts. The real win turned out to be reproducibility — stochastic proposals can't be A/B tested, and temperature 0 doesn't fix that;
Semantic recall still needs an embedder (local or external), stamped into store metadata to catch mismatches; CI fails if published figures drift;
Limits stated up front: improves memory, never model weights; no daemon.

Repo: github.com/AreevAI/areev

Original post →

Agent memory as one SQLite file: 30µs recall, model-free self-improvement loop

More from coding & agent