Tencent open-sources T-Mem, an associative-recall memory architecture for long-term AI memory

aigclink · x · 2026-09-21

Tencent has open-sourced T-Mem, a memory architecture targeting the overlooked weakness of associative recall in long-term AI memory. Instead of just storing "what something is," T-Mem has the LLM anticipate, at write time, which seemingly unrelated future queries a memory might serve, indexing those associative cues—shifting memory from passive retrieval to active anticipation, key for companion AI and digital humans.

Design: four triggers in a 2×2 matrix—two evidence granularities (atomic Item / full Scene) × two recall types (descriptive / associative).

Benchmarks:

80.26% accuracy on standard LoCoMo, 74.81% on LoCoMo-Plus (−5.45 pp)
HyperMem drops from 77.01% to 48.63% (−28.38 pp), showing T-Mem's much stronger robustness on associative recall.

Original post →

Tencent open-sources T-Mem, an associative-recall memory architecture for long-term AI memory

More from coding & agent