AI Dictionary › AI Fundamentals

Context Compaction

Compattazione del contesto

Context compaction is the operation by which an AI system, as a conversation approaches the limit of its context window, summarizes what has happened so far and replaces the full text with the summary. The agent keeps working with a shorter but still coherent memory, instead of stopping or forgetting everything. It is by now a standard feature of assistants that work for hours on the same task.

Definition

It has two opposite effects on drift. It cuts the thread by which the model imitates its own earlier text, and so it breaks the habits built up in a long session. But it also cuts the corrections received: an instruction given at the start and not carried into the summary disappears, and the model returns to its initial behavior without knowing it. What survives compaction depends on who writes the summary, that is, the model itself.

The operating rule: instructions that must survive are not left in the conversation, they go where compaction cannot reach: in the system prompt, in a file re-read at every step, in an external note.

Related terms

• Context Window
• Drift
• Context Engineering
• Agentic AI
• Context Length

More in AI Fundamentals

• LLM (Large Language Model)
• Token
• Context Window
• Temperature
• AI Hallucination
• Fine-tuning
• RAG (Retrieval Augmented Generation)
• Embedding

Put it into practice

• Real scenario: apply this concept with an instant AI verdict.
• Daily gym: short drills to train dictionary concepts.
• Playoff game: challenge other players, elimination-style.

Say what you mean. Get what you need.

Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.