AIT-Mem’s trigger-based approach to AI memory recall
TencentAI_News describes T-Mem as an open-source architecture that retrieves memories by situation, even without keyword overlap, using triggers attached when memories are stored.
TLDR
TencentAI_News says T-Mem stores memories with triggers that anticipate when they might matter again, then uses those triggers to retrieve memories at query time. The account says the architecture works with any LLM without fine-tuning or model changes. On LoCoMo-Plus, described as stripping keyword overlap to test associative recall, it reports drops of 28 to 50 points for mainstream systems versus 5.45 points for T-Mem. The post links a paper and MIT-licensed code.
