AI agents that remember everything have a problem: old facts can override new ones, or vanish when they are still useful.
A paper posted to arXiv (arXiv:2609.10263) describes RD-Forget, a training-free framework that separates what an AI agent stores from what it actually uses to answer a question. The system keeps every observation in a permanent archive, then builds a separate, query-specific memory view for each answer. A frozen language-model curator sorts facts into semantic slots and flags when a new fact replaces an old one, so outdated values get suppressed for present-tense questions but stay retrievable if someone asks about the past. The paper reports tests across conversational memory, knowledge updates, fact consolidation, long-context reasoning, and personalization tasks, all run through a shared answering pipeline, with configurations lacking the forgetting step or query-conditioning step scoring worst.
The idea reframes a persistent headache for chatbots and AI assistants: agent memory has mostly meant bigger context windows or vector databases, not a mechanism for deciding when an old fact should stop counting. RD-Forget's slot-based, retrieval-time suppression looks like a direct answer to the classic failure mode of an assistant confidently repeating a stale detail, like an old phone number, because nothing ever told it that detail expired.
Training-free is the operative word here. If the ablation results in arXiv:2609.10263 hold up under independent testing, this could be a cheap add-on for existing memory systems rather than a reason to retrain anything.