The authors separate two demands on an agent's memory: evidence access, keeping the relevant steps intact, and aggregation, combining events into counts, lists or timelines.
Agent memory benefits more from event aggregation than retrieval alone
TallyMem, a system that executes model-written event specifications over a reusable ledger, raises exact match by 13.8 points on held-out MINTEval repository histories.
Chinese Tech
Zerui Yang · Quan Xue · Chaorui Zhang · Xueyan Niu · Xuanyu Zheng · Xiyuan Wang · +5 more
Huawei Technologies Ltd.
Research Digest··2 min read
The authors diagnose two distinct bottlenecks in agent memory, access and aggregation, and show that retrieval alone fails when answers require combining many events.
Why this paper
From Huawei Technologies Ltd.
In one line
TallyMem improves exact match on held-out MINTEval repositories by 13.8 points over Trajectory Memory.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§