Context Window Compaction: What Gets Kept and Lost

Context window compaction replaces some conversation history with a shorter representation, often a summary, so an AI can continue within its working capacity. The summary can preserve the main task while omitting details that later matter. fenbs keeps reviewed decisions and progress outside that summary, ready for the next session to read.

A context window is the material available to an AI for its current response. It can include instructions, messages, document excerpts, and tool results. Compaction changes that material; it does not turn a conversation into perfect permanent memory or guarantee that the next answer will reflect everything said earlier.

Why does a conversation need compaction?

A long task accumulates material. Your original request is joined by follow-up questions, generated drafts, source files, and checks. Eventually, keeping every detail competes with reading the next document or producing the next answer. Summarizing earlier discussion creates room for that continuing work.

Anthropic describes its API compaction feature↗ (opens in a new tab) as replacing older turns with a server-written summary. That is one product implementation, not a promise that every AI app uses the same algorithm, trigger, or summary format. Check the app’s documentation before assuming identical behavior.

For example, a designer may start with a landing page brief, explore several layouts, and then revise the chosen layout. The abandoned drafts are less useful than the final choice, but the reason a certain color was rejected might still constrain the work. The difficult part is choosing which details remain important.

For the underlying capacity concept, see AI context windows. This article focuses on the change from detailed conversation to compacted context rather than on the maximum size of a particular model.

What gets kept, shortened, or left out?

There is no universal preservation list. A summary may keep goals, recent progress, and important conclusions while dropping exact wording and intermediate output. Treat the resulting record as a useful interpretation of earlier work, then inspect the product-specific behavior for instructions and source material.

Suppose a developer investigated three possible causes of a bug and disproved two. A compacted note saying “investigating login failure” invites repeated work. A useful checkpoint says which explanation remains plausible, what evidence ruled out the others, and which conclusion is still uncertain.

The aim is to preserve distinctions, not every sentence. Put the information the next action depends on in a record you can review, such as fenbs Where we left off.

Is compaction the same as compression or retrieval?

People sometimes use “context window compression” as a broad label for making input smaller. Compaction through summarization is usually a meaning-based reduction: it rewrites content into a shorter account. Do not assume you can reverse it and reconstruct the original wording, as you could with an ordinary compressed file.

Retrieval is a different operation: find relevant material from a larger collection and bring selected parts into the current task. A workflow can use both. It might summarize earlier discussion while separately reading the source document needed to answer today’s question. A remembered summary should not replace a source check when exact wording matters.

Caching also has a different job. It can reuse processing for repeated input; it does not by itself create a reviewed handoff or tell you which decision is current. A change in speed, usage, or visible history is not enough to diagnose compaction without the app’s own indicators.

These distinctions matter when you investigate AI context rot, where the available material becomes less useful. Compaction can reduce clutter, but a summary that retains outdated assumptions can also carry those assumptions forward.

How can you protect decisions before context changes?

1. State the current goal and the boundaries of the work in plain language. Include what would make the result acceptable.
2. Separate approved decisions from suggestions. Name the person who chose and the reason when that information matters.
3. Record evidence for completed work and label unverified claims. Keep source links or file paths beside important conclusions.
4. Write the next action and unresolved questions. Review the record before clearing, compacting, or handing off the session.

Keep a stable project rule distinct from a temporary instruction. “Use the approved vocabulary” may apply across sessions; “check the second draft next” expires when that check finishes. Without that distinction, a short summary can present old progress as a standing instruction or bury a lasting rule among temporary details.

In fenbs, Decisions and rules and task progress have separate places. That helps an assistant read the current project context without treating every earlier comment as equally authoritative.

What should you verify after a summary replaces history?

Ask the assistant to restate the goal, restrictions, known evidence, and proposed next action. Compare its answer with your maintained record. Correct anything missing before asking it to make a change, send a message, or draw a conclusion that depends on the omitted detail.

For coding work, inspect the current files rather than assuming the summary describes their latest state. Anthropic’s Claude Code context guide↗ (opens in a new tab) distinguishes material reloaded from disk from conversation material summarized away. Product behavior matters more than a generic expectation that everything “comes back.”

Use a specific probe. Ask which requirement remains untested or whether a particular action needs approval. “Do you remember everything?” cannot establish either fact. If you find a gap, restore the minimum relevant context and verify the affected plan rather than repeating the entire project history.

A repeatable routine for selecting future reads belongs in context window management. Compaction is one event in that broader practice, not a substitute for choosing useful inputs throughout the task.

How to do this with fenbs

1. Create a free account, add the project, and put the active work into a task.
2. Record approved constraints in Decisions and rules and keep unresolved choices open for a person to decide.
3. Save the task’s completed checks, remaining work, and Where we left off before a major context change.
4. Connect your AI app and ask it to read project context before continuing. Confirm the summary of its next action against the saved record.

fenbs does not expand a model’s context window or recover missing conversation text. It fits work that needs a maintained record across sessions and assistants; a single short question may need no extra system. One person can start free, with no credit card.