More setups are coming soon. This page will be updated as their results land.
Hindsight retains every session of a long conversation, then answers each question with reflect. A judge compares each answer with the expected one. Questions LoCoMo gets wrong (an answer the conversation never states, or the wrong person named) are removed.
Results from: Agent Memory Benchmark
At commit: 28b644b
Dataset: LoCoMo, 10 conversations, 1502 questions
Judge: gemini-3.1-flash-lite
| Memory | ||||||
|---|---|---|---|---|---|---|
 Hindsight gemma-4-26b-a4b | 85.5 | 67.7% | 89.2% | 81.5% | 84.5% | 10.6 s |