How working context is assembled
Clarify the need
The current question determines which rules, facts and earlier decisions are relevant.
With LeaWhat is needed: Lea’s preference and this week’s free slots. Old calendars do not help.
Fetch relevant content
Search and file access supply the required excerpts from documents, tools or stored knowledge.
With LeaA calendar search returns the free slots. The note about Lea comes from memory.
Assemble the working context
Selected content is passed to the model together with the task. Source references make later checks easier.
With LeaTask, rules, note and calendar results go to the model together.
Update and condense
New results extend the context. In long runs, summaries can free up space.
With LeaA summary replaces older calendar details. The reference to the calendar stays.
TRY IT YOURSELF
Pack the context window.
The window holds 100 units. Choose what the model needs for the slot with Lea and watch how the answer changes.
Context window0 / 100
Striped: 15 units reserved for output. The answer needs room in the window too. All numbers are schematic.
How the model responds
Does not fit the window.
No room remains for the answer. The application would have to trim or condense. Take something out.
No task in context.
The model sees material but no task. The answer can only guess what you want.
The free slots are missing.
The model knows the task but not the calendar. It can propose a tool call or ask, but cannot name a slot yet.
Lea’s preference is missing.
Proposal: Thursday, 3 pm. Formally free, but Lea prefers mornings. It was stored, but it was not in context.
The rules are missing.
The proposal fits. But the model does not know your instructions for this run. The runtime has to enforce them either way.
Fits, but costs space.
Proposal: Thursday, 9:30 am. The extra material does not help with this task and takes up space later tool results will need.
Exactly what is needed.
Proposal: Thursday, 9:30 am. The window still has room for tool results and the answer.
TRUE OR FALSE?
Three things people often say.
A larger context window automatically makes answers better.
Correct.Not quite. False It extends capacity. Quality still depends on the selection and clarity of the content.
If an agent can access a repository, everything is in context.
Correct.Not quite. False The repository can be far larger than the active context. Search and targeted reading bring in relevant parts.
Condensing can lose details.
Correct.Not quite. True That is why decisions, open questions and source references should remain. Original sources allow later checks.
0 of 3 correct.
Explore the mechanisms
A bounded workspace
The context window limits how many tokens a model call can process. Tokens are small processing units, such as word fragments. Output also requires space. A larger window increases capacity; quality still depends on the selection and clarity of its contents.
Selection makes access useful
An accessible repository or document archive can be much larger than active context. Search and selective reading bring in relevant parts. Stored memories also become working material for the current call through loading or retrieval.
Assess matches with provenance
Retrieval means fetching relevant information. Semantic search can find similar meanings despite different wording. Source references and document metadata then help assess a match. Similarity alone does not establish its freshness or factual accuracy.
Compaction preserves a selection
Compaction replaces older conversation portions with a shorter representation. This supports continuation but can lose details. Important decisions, open questions and source references should remain; original sources enable later verification. Persistently stored memory is a separate storage area.
Fundamentals and primary sources reviewed on 5 November 2026. The examples explain the principle; specific products may differ.