A long‑running assistant cannot retain every observation, so it compresses its history into summaries. Summarization is not neutral: what is kept is chosen based on an assumed purpose of the record, and that choice is made before any of the user’s standing goals actually query the memory. Goals usually agree on what happened; they differ on which parts are worth the limited space. Once the history becomes too long to reread, the summary replaces the stream and omitted details are lost forever.
We investigate how a memory should summarize when it serves several standing goals simultaneously. Three write‑time policies are compared: (1) summarize without any goal in view, (2) write a single summary that tries to cover every goal, and (3) write a separate per‑goal summary and read them together. Experiments span multiple models and event streams, keeping the read step fixed and varying only the write step.
Findings show that summaries written for different goals overlap each other less than a summary overlaps its own rewrite. Per‑goal summaries outperform the others in relevance, completeness, and accuracy, while the all‑goal summary performs worse than even the neutral one despite using a larger budget. The takeaway is that interpreting at write time pays off, but only for the goal that later asks.
Review