T2Mem is a framework that defers memory construction to the inference phase, building it progressively along the reasoning process. It uses lightweight offline metadata as retrieval cues to reconstruct dependencies across distant interaction sessions.
HOW THIS AFFECTS YOU
●
builderYou can reduce memory management overhead by building context adaptively at test time rather than pre-computing it.
●
researcherThis approach allows for more query-specific memory construction, reducing noise in long-context retrieval.