The authors address two weaknesses in existing long-term memory systems for LLM agents: temporal locality (related messages cluster in short episodes, even when surface similarity to a query is low) and temporal evolution (messages within an episode form a state trajectory, where proposals are revised, superseded, or supported).
Temporal structure improves long-term memory retrieval for multi-party LLM agents
T-Mem recovers locally concentrated evidence and reconstructs state transitions to outperform flat retrieval and existing agent-memory baselines.
Chinese Tech
Junyou Guo · Chongyang Tao · Yan He · Yong Wang · Shuai Ma
Beihang University · Alibaba Group
Research Digest··2 min read
Guo et al.
Why this paper
From Alibaba Group and Beihang University
In one line
Temporal structure, used for evidence recovery and state-transition reconstruction, improves long-term multi-party conversational 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
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§