TrackMem: Tracking Evolving Beliefs in Long-Horizon Financial Dialogue
TrackMem uses per-speaker graph memory to track revised figures and shifting stances across sessions in financial dialogues where facts change over time.
Existing memory systems fail on long-horizon financial conversations because figures are revised and stances shift without restatement. TrackMem maintains a graph that separates perspectives by entity and lens while preserving chronological speaker actions. It outperforms baselines on FinLoCoMo benchmark reaching 73.9% macro accuracy versus 69.4% for the strongest baseline especially on questions about revised figures.
- 73.9% accuracy on failure-mode split with 765 questions
- 86.5% accuracy on revised figures questions
- Separates perspectives by entity and analytical lens
- Preserves speaker actions in chronological order
From the original article
1 min read at arxiv.orgMemory systems for LLM agents are evaluated mostly on social dialogue, where facts accumulate and rarely change, so remembering reduces largely to finding the turn in which something was said. Financial dialogue breaks this assumption: figures are revised, stances shift without being restated, rules stated once govern later questions, and a single company is discussed under several analytical lenses.
By Austin M. Y. Cheung · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.