Summarization Memory
Also called Memory Compression
Squeezing old conversation into a short recap so a long chat still fits in the window.
Think of it like
Replacing a two-hour meeting with a five-line minutes doc you can skim later.
Example
After 40 messages, the app quietly compresses the early ones into "User is planning a Japan trip in October, budget-conscious, vegetarian."
How it actually works
When history won’t fit, you summarize the oldest turns into a compact note and keep only that plus recent messages. It buys room but leaks detail — the summary decides what survives, and anything it drops is effectively forgotten. Good systems summarize hierarchically and keep raw text retrievable in case the recap missed something.
For product teams
Lets long-running assistants feel continuous, at the cost of occasionally forgetting a detail the summary skipped.
For engineers
Recursive summarization of stale turns; lossy compression traded against context budget, ideally paired with retrievable raw logs.
Related
- Conversation History — Compresses the conversation history.
- Context window — A workaround for the context window limit.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome