A new arXiv paper argues that long-running AI assistants need a different memory summary for every goal they track, not one all-purpose version.
The researchers tested three ways an assistant can summarize its event history: write a single neutral summary with no goal in mind, write one summary meant to cover every goal at once, or write a separate summary per goal and read them together at query time. They held the retrieval step constant across several models and event streams, so only the summarization strategy varied. The goal-specific summaries overlapped each other less than a summary overlapped a rewrite of itself, meaning different goals really do want different information kept. Per-goal summaries won on relevance, completeness, and accuracy.
That matters because most assistant memory systems today compress history once, early, before anyone knows which future question will need it. This study suggests that approach quietly throws away details that matter to specific tasks, even when the summary looks reasonable on its own. The more surprising result: the single all-goal summary performed worse than the goal-blind one, despite having far more context to draw from.
Bigger context windows will not fix this on their own if the write step still guesses wrong about what to keep.