A Better Newspaper

A set of 2026 papers addresses how AI agents should manage, retrieve, and retire long-term memory under limited context budgets — a growing area as agents are deployed for extended, multi-session interactions. ## Key Contributions - **Evidence composition over retrieval**: conventional memory access retrieves and scores memories individually, often wasting context slots on redundant facts while missing complementary evidence; one paper reframes memory reading as query-conditioned evidence composition, valuing each memory by what it adds beyond already-selected evidence (arXiv:2610.15642, October 14). - **Factorized skill memory**: studies forgetting in continual skill internalization, finding a hierarchical structure to how agents lose the ability to invoke previously learned procedural skills as new skills are learned (arXiv:2610.15492, October 14). - **Memory-retirement gating**: proposes a three-way task contract (SUPERSEDED vs. NEEDS-CONFIRMATION vs. COMPATIBLE) for deciding when a new utterance should retire an old stored claim, evaluated against a labeled 500-pair test set (arXiv:2610.15710, October 14). - **Long-context via sparse episodic memory**: challenges the assumption that a linear-attention model's fixed-size recurrent state must bear the full burden of long-term storage, introducing NoteBank, a sparse episodic memory that archives selected text for later retrieval (arXiv:2610.15715, October 14). ## Strategic Significance Memory management is a core bottleneck for deploying persistent AI assistants in enterprise and consumer contexts (e.g., customer service, personal assistants, clinical documentation). How agents decide what to remember, forget, or retire has direct implications for data accuracy, liability (e.g., acting on stale information), and compliance with data retention/deletion obligations. This is likely to remain an active and commercially important research area.