T-Mem: Memory That Anticipates, Not Archives

In three linesT-Mem proposes a long-term conversational memory architecture that overcomes lexical and vector similarity bounds. The system introduces write-time triggers to enable two recall modes: descriptive (surface features) and associative (latent semantic arcs). T-Mem achieves state-of-the-art on LoCoMo and LoCoMo-Plus benchmarks.

Summary generated by Claude — human-verified