What Is AI Context Compaction? (2026): Shorten the Window, Keep the Pin

AI context compaction is the step that shortens material already in the window: prior turns, tool results, or retrieved passages, while a pinned instruction stays whole. The next model call receives the short form plus the pin. Shrinking the prompt template itself is what prompt compaction is. The failure when nothing fits is what AI context window overflow is. A block that must not be shortened is what AI context pinning is.

The limit you are trying to fit is what a context window is. This page does not invent a token count for any model. If a product prints a limit, use that product’s page. Do not copy a number from memory into this definition.

Claude’s prices are from claude.com/pricing on October 24, 2026. ChatGPT Plus is $20 a month on learn.chatgpt.com/docs/pricing that day. Gemini and Notion prices are from directory pages checked October 16, 2026. Published October 24, 2026. We did not measure truncation. The directory does not list a compaction fee.

What may be shortened

Block	Compaction may	Compaction may not
Prior turns	Replace them with a short form you can audit	Drop a turn the pin says to keep verbatim
Tool results	Keep the fields you named	Keep a secret the redaction pass should have masked
Retrieved passages	Keep the ids and the lines you marked	Replace a passage with a paraphrase and still cite the id as if the words were original
Pinned instructions	Leave them whole	Summarize the refusal rule to save space

flowchart TD
  full["The window holds the pin, prior turns, tool results, and passages"] --> pin["The pin is copied unchanged"]
  full --> short["The other blocks are shortened by a rule you wrote"]
  pin --> next["The next call receives the pin plus the short form"]
  short --> next
  next --> lost{"Did a required id or a required sentence disappear?"}
  lost -->|Yes| undo["Do not send that call. Fix the rule"]
  lost -->|No| go["The call may run. Overflow is a separate failure"]

A summary written for a person is a document they will read. Context compaction is an input to a machine. If the machine loses the constraint, the shorter window failed, even if a person would have understood the summary.

Where the rule is kept

The rule. Claude Pro is $17 a month billed annually, $200 up front, or $20 monthly. The Project holds what may be shortened and what the pin contains. The review is Claude review.

One draft of the rule from a paste. ChatGPT Plus is $20 a month. Discard a version that summarizes the pin. The ladder is ChatGPT pricing explained.

A long doc that already lives in Google’s tools, when you are deciding what to pin. Gemini Google AI Plus is $4.99 a month on the October 16 card. Google AI Pro is $19.99 a month on that card. The doc is the source. Compaction is not a reason to skip opening it. The ladder is Gemini pricing explained.

The register of pins. Notion AI Business is $20 a seat a month. One row per workflow: what is pinned, what may be shortened, who owns the rule. The invoice is Notion AI pricing explained.

A prompt that holds the line:

“Draft the context-compaction rule from the blocks I said may be shortened and the pin I pasted. Leave the pin verbatim. If I did not name a pin, write Pin not set. Do not summarize the pin.”

What not to delegate

Do not ask the model to compact “whatever does not fit” with no pin. Do not paraphrase a cited passage and keep the citation. Do not store the full secret-bearing tool result in the short form because the summary “needed detail.” Mask first. The pass is what an AI output redaction pass is when the text is about to leave. The same list applies to what you keep inside the window.

FAQ

How is this different from prompt compaction?

Prompt compaction shrinks the prompt template so the task and the instructions still fit a budget. Context compaction shrinks history that accumulated after the template: turns, tools, passages. You can compact the prompt and still overflow on the transcript. You can compact the transcript and still need the template shortened. They are two steps.

How is this different from pinning?

Pinning names a block that stays byte-for-byte. Compaction is what you do to the rest. A pin with no compaction rule just sits at the top until the window overflows. A compaction rule with no pin will shorten the instruction you meant to keep.

How is this different from overflow?

Overflow is the failure: the request no longer fits, and the product drops text, summarizes it, or errors. Compaction is a step you run on purpose before that failure, with a rule. If you did not choose the rule, the product’s overflow behavior is not compaction. It is a drop you did not audit.

Who may change what the pin contains?

The owner of the workflow. A shorter pin that drops a refusal is a product change, not a savings. Record the version. The next call uses the new pin only after that owner accepts it.

Last verified

October 24, 2026 for Claude and for ChatGPT’s plan page. October 16, 2026 for Gemini and Notion. Published October 24, 2026.

Sources: Claude pricing, ChatGPT and Codex pricing, Gemini, Notion pricing.

Published by Andrew for AI Tools Optimizer.

Claude is in the chatbots and assistants directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.