Conversation summarization helps an assistant or agent carry useful information forward when a chat becomes too long. A good summary prioritizes stable facts, user preferences, decisions, constraints and open tasks rather than trying to preserve every sentence.
A running summary for a project assistant might hold the agreed goal, decisions made, constraints, open questions and the user's stated preferences. Update it every few turns rather than only when the limit is hit. Keep recent messages verbatim and summarize older ones, so the model still sees exact wording where it matters most within the context window.
Summaries can introduce omissions or distortions, so keep critical instructions and source records available when accuracy matters. Mark uncertainty and update the summary when the user corrects a previous detail. Related: Context Compression, AI Agent Memory.