LoCoMo

LoCoMo measures multi-session conversational memory across 10 dialogues and about 300 questions. It comes from mem0ai/memory-benchmarks.

• Single-hop: direct factual recall from prior turns.
• Multi-hop: joins facts across sessions.
• Temporal: reasons about order, recency, and time.
• Open-domain: mixes memory retrieval with broader answering.

Reference values come from published materials. Hindsight per-type LoCoMo breakdown is not shown here. GBrain has no public LoCoMo run.

Quaid is doc-native today. The current adapter stores whole conversation turns as documents instead of extracting and maintaining distilled facts. That limits multi-hop and temporal performance on the first run.

The roadmap item that closes this gap is the conversation memory feature in issue #105. Once Quaid can ingest turns as durable memory facts, this benchmark becomes a direct before-and-after measurement.