BLOG

Field notes

What we build, what we measure and what we learn running AI assistants for real work.

Benchmarks · 23 October 2026

NeoCore, measured

Open-source memory for any AI harness, tested: NeoCore on LongMemEval-S and LoCoMo, and its 32M NeoSift filter against open rerankers, a hosted filter and frontier LLM judges.

Read the post ↗NeoCore 0.2 · 22 October 2026

Fewer, better memories

NeoCore 0.2 adds NeoSift: a 32M-parameter model that runs on your graphics card and decides which recalled memories would actually help with this prompt.

Read the post ↗