Debate: Should LLM memory focus on forgetting and updating, not recall?

A Reddit discussion argues long-context models that handle hundreds of thousands to millions of tokens still often reuse outdated facts when old and new information coexist. Using an API timeout changed from 30 to 5 seconds as an example, it examines update mechanisms like DeltaNet and BDH and questions whether long-context benchmarks test answer revision under contradictory evidence.