Long-term conversational memory has a bandwidth problem: a model needs enough evidence to answer correctly, but feeding it everything makes the context unwieldy. A new arXiv preprint, JustMem, frames this as a need for "just-enough memory access" — retrieving sufficient evidence without indiscriminately expanding the context presented to the language model.
The paper highlights the core difficulty: relevant evidence may be distributed across a long conversation, so it cannot be captured by a single obvious retrieval step. The motivating idea is that the right amount of context is not necessarily the maximum amount.
Because this digest draws on only one source, there are no differing findings to note. The abstract is brief, so the main contribution here is the problem framing rather than a detailed account of the method or its evaluation.