All-Mem: Agentic Lifelong Memory via Dynamic Topology Evolution

ArXi:2603.19595v1 Announce Type: cross Lifelong interactive agents are expected to assist users over months or years, which requires continually writing long term memories while retrieving the right evidence for each new query under fixed context and latency budgets. Existing memory systems often degrade as histories grow, yielding redundant, outdated, or noisy retrieved contexts.