Large language models that read long documents often use a “memory agent”: instead of keeping the whole document in context, the model reads it piece by piece and keeps a short summary (a fixed-size memory buffer) updated as it goes. A recent method called MemAgent does this well, but has one weakness: every time it updates the summary, it rewrites the entire buffer from scratch. It has no way to check which old information is still needed before throwing it away, so useful facts can g et lost as the document gets longer. This paper proposes MemAgent+CITRN, which adds a way to make that decision on purpose instead of by accident. Three simple ideas make this work. First, a small side-network learns to score each piece of memory by asking, in effect, “how much would the model’s expected accuracy drop if this were deleted?” Second, instead of a hard keep-or-delete rule, deletion is a soft, trainable probability, so the model can learn good deletion habits directly from data. Third, deleted information is not thrown away immediately, it is kept in a smaller backup store and can be brought back later if it turns out to be needed after all. We describe the full design, explain every part with precise definitions, and prove several mathematical guarantees about when this approach should reduce mistakes, being careful to state exactly what is and is not proven. We also compare the design against very recent competing methods and lay out, in detail, the experiments needed to test whether it works in practice, since those experiments have not yet been run.

Long-context language models,reinforcement learning,Memory Management,Counterfactual reasoning,selective retention,token relevance estimation