What is context editing?
Context editing is the automatic removal of stale content, such as old tool results or earlier reasoning blocks, from an agent's context window as it grows, so a long run keeps going without summarising the whole history.
Most of an agent's context is tool output it has already acted on: search results it has read, files it has already edited, query rows it has already summarised. Context editing removes that material once it is stale, instead of summarising everything or stopping when the window is full.
Anthropic introduced context editing on its developer platform in September 2025, alongside a file-based memory tool. On an internal agentic search eval, context editing alone improved performance by 29%, and by 39% combined with the memory tool. On a 100-turn web search eval it let agents finish workflows that otherwise failed from context exhaustion, while cutting token use by 84%.
How context editing works
Tool result clearing
The oldest tool results are replaced with a placeholder once the context passes a threshold. Tool call inputs can optionally be cleared too.
Thresholds and limits
In Anthropic's API, clearing triggers at 100,000 input tokens by default and keeps the three most recent tool uses; both are configurable, and named tools can be excluded.
Reasoning block clearing
A second strategy removes thinking blocks from earlier assistant turns, keeping a set number of recent ones.
Applied on the server
Edits are applied before the prompt reaches the model, while the client keeps the full, unedited history.
Context editing, compaction and memory
Compaction summarises the history and replaces it; context editing deletes specific content and leaves the rest untouched. Anthropic's documentation now treats server-side compaction as the primary strategy for most long conversations, and context editing as the option for finer control over what is removed.
Context editing pairs naturally with memory. When the context nears the clearing threshold, the agent can be warned and save anything it still needs to memory files before the tool results disappear, then read them back later.
The prompt caching trade-off
Clearing content changes the prompt, so it invalidates the cached prefix from that point. Anthropic's API offers a minimum amount to clear on each pass, so that every clear removes enough tokens to justify the cost of writing a new cache entry.
How Motiontic's engineers apply it
Motiontic's engineers set clearing thresholds from eval runs, exclude tools whose results stay relevant for the whole task, and give the agent a memory file for facts it must keep. The aim is a context that holds what the agent needs next, not everything it has seen.
Context editing compared with compaction
| Context editing | Compaction | |
|---|---|---|
| What happens to old content | Specific items are cleared | The history is summarised and replaced |
| What survives | Everything not cleared, word for word | A summary of what came before |
| Control | Fine: which tools, how many to keep, when to trigger | Coarse: when to summarise and what to stress |
| Typical use | Tool-heavy runs with bulky, short-lived results | Long conversations that need continuity |
| Cache impact | Invalidates the cache from the cleared point | Replaces the prompt after each summary |
Common questions
Does context editing lose information?
Yes, from the model's view: cleared results are gone from its context, though the client keeps the full history. Excluding important tools and saving key facts to memory first limits the loss.
What is the difference between context editing and compaction?
Context editing removes specific items, such as old tool results, and leaves everything else as it was. Compaction rewrites the history into a summary.
Does context editing break prompt caching?
It invalidates the cache from the point where content was cleared. Clearing in larger batches, rather than a little at every turn, keeps the cost of new cache writes down.
Related
Compaction
Summarise the old turns, keep working: the standard move for long agent runs.
Context engineering
Curating everything in the model's view, not just the wording of the prompt.
Context rot
Why agents get worse the longer they run, and what keeps context clean.
Agent memory
Session state for the task at hand, plus stored knowledge the agent recalls later.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.