Evals

turbomem publishes a frozen eval snapshot so customers can inspect how memory quality is measured, not just a marketing claim.

Numbers live on turbomem.dev/evals. The snapshot file is packages/evals/results/latest.json in the library repo.

Suite	What it measures	Headline metric
Product	Extraction, semantic retrieval, dedup, scope isolation, local latency	recall@5, leak rate
LoCoMo	Multi-session conversational QA after add + search	official token F1

This page is the index. Read next:

• Methodology - pipeline, models, k, what we do not claim
• LoCoMo protocol - dataset IDs, scoring, ingest
• Product goldens - dataset cards
• Running evals - commands, cost, CI