# What is Conversation Summarization?

Also called summary memory.

Conversation summarization compresses earlier turns of a dialogue into a shorter running summary so a session can continue past the limits of what fits in a single prompt. The summary replaces the omitted turns, usually while the most recent messages are kept verbatim. Every summarization pass discards detail permanently.

The standard pattern keeps a sliding window of recent turns in full and a rolling summary of everything before it. When the window overflows, the oldest turns are folded into the summary and dropped. This bounds prompt size regardless of session length, which is what makes long-running assistant sessions possible at all, and it is why the technique is sometimes called summary buffer memory.

Loss is inherent and cumulative. Each pass summarizes a text that was already a summary, so specifics degrade with every fold: exact numbers become approximate, named constraints become general ones, and decisions lose the reasoning behind them. Summaries that are re-summarized many times tend toward vague narrative, which is why some systems summarize from the original transcript each time rather than from the previous summary.

What to preserve should be explicit rather than left to a generic instruction. Prompts that name the categories to retain, such as decisions made, constraints stated, identifiers mentioned, open questions, and user preferences, produce far more useful summaries than an instruction to summarize the conversation. Numbers, names, and file paths deserve a rule to copy them exactly.

Summarization pairs naturally with durable memory. Facts worth keeping beyond the session should be extracted into a separate store rather than left inside a summary that will be compressed again, and the raw transcript should be retained where policy allows so a user can be shown what was actually said. Summarization is a context management technique, not a memory system on its own.

## Key points

- Folds older turns into a rolling summary, keeps recent turns verbatim
- Bounds prompt size for arbitrarily long sessions
- Detail loss compounds each time a summary is re-summarized
- Name the categories to preserve, including exact numbers and identifiers
- Durable facts belong in a separate store, not in the summary

## In practice

After ninety turns of debugging, the prompt carries the last twelve messages verbatim plus a summary noting the stack, three ruled-out causes, the exact error string, and the agreed next step. A generic earlier summary had said the user is debugging a deployment issue, which lost the error string and caused the assistant to re-suggest a fix that had already failed.

## Related terms

- [Short Term Memory](/en/glossary/short-term-memory)
- [Working Memory](/en/glossary/working-memory)
- [Memory Consolidation](/en/glossary/memory-consolidation)
- [Long Term Memory](/en/glossary/long-term-memory)
- [Context Injection](/en/glossary/context-injection)

[Back to the AI Glossary](/en/glossary)