Tiger Data Reports F1 0.665 on LoCoMo with a Single Postgres Table
Summary
Tiger Data's Matvey Arye reports that storing raw conversation turns in one Postgres table and giving an agent three MCP search tools scores F1 0.665 on LoCoMo with Claude Sonnet, above Omni-SimpleMem's reported 0.613. Dropping LLM-extracted atomic facts did not lower F1 and cut about 50% of ingestion time, the company says.
Key Points
- Tiger Data says removing the six-call cap lifts Sonnet F1 from 0.615 to 0.646, a paired gain of 0.031.
- Arye reports a nine-parameter search schema costs 0.09 F1 versus a one-parameter tool on the tuning conversation.
- Adversarial F1 falls from 0.922 to 0.870 across development while multi-hop rises 0.105, Arye reports.