LoCoMo

Dated snapshot — verified 2026-08-27

Long-term conversational memory. Built by a machine-human hybrid pipeline: two LLM-driven personas hold multi-session dialogues grounded on persona profiles and temporal event graphs. Conversations span up to ~32 sessions, ~600 turns, ~16K tokens on average. ~1,540 questions across single-hop, multi-hop, open-domain and temporal categories.

What it measures well

Multi-session recall over a realistic conversational shape — the same shape as this course's canonical corpus, at larger scale.

What it misses

Knowledge updates and abstention. LoCoMo does not systematically test what happens when a fact changes, which is the failure this entire course is built around. A system that never supersedes anything can score respectably here. Do not read a LoCoMo number as evidence about staleness handling.

Reading its numbers

Widely cited and widely re-run under different conditions — backing model, retrieval budget, and which subset was used all move the result substantially. Numbers are comparable only within a single report.