Loading the context…
Loading the context…
Condensing earlier conversation or working material so a model can continue with a smaller retained context.
The system creates a summary and replaces some earlier context with it, sometimes retaining recent messages or links to files. Choosing what to keep matters: a shorter record can omit a constraint or the reasoning behind a decision. Durable notes and retrievable original material can complement the summary. Summarization itself also requires processing, so compaction is not automatically free or instantaneous.
Suppose an assistant has 30,000 tokens of conversation about preparing a budget. A 3,000-token summary preserves the approved J$2 million ceiling, the chosen supplier, pending quotations and the next calculation. That removes 27,000 tokens from the carried-forward history. If the summary omits the instruction that delivery must occur before July, the assistant may propose an unsuitable purchase despite retaining the headline budget. The example illustrates both the space saving and the loss risk.