Your agent's context window is a budget, not a diary — here's how Anthropic spends it

A long agent session doesn't run out of intelligence. It runs out of room. Every tool call, every file it reads, every retry gets added to the transcript, and none of it leaves on its own — until the window is mostly the record of what the agent already tried, with less and less space for what it's supposed to do next. Anthropic's engineering team has published a specific answer to this, and it is less "write a better prompt" and more "manage a shrinking budget."

The reframe: context engineering, not prompt engineering

Anthropic's own framing is a single sentence: find the smallest set of high-signal tokens that maximize the likelihood of your desired outcome. That's a different goal from "fit as much relevant material as possible." A long agent session accumulates information whether or not it's still useful, and the fix isn't a bigger window — it's actively deciding what stays.

Context editing: clear what's stale

The most direct tool is context editing, which automatically removes old tool calls and their results as a session approaches its token limit, rather than leaving them to sit as dead weight. In Anthropic's internal evals, this alone produced a 29% improvement on an agentic search task. Combined with the memory tool below, that rose to 39%. In a separate 100-turn web-search evaluation, context editing cut total token consumption by 84% — and let the agent finish workflows that failed outright without it, because the transcript ran out of room before the task did.

Memory that lives outside the window

The memory tool gives an agent a directory it can create, read, update, and delete files in, and that directory persists across conversations — it isn't part of the context window at all. Instead of re-deriving what it already figured out three turns ago from a transcript that keeps getting trimmed, the agent can write it down once, outside the conversation, and read it back when needed.

Compaction: summarize, then start over

For sessions too long to keep patching, compaction summarizes the history and reinitializes the agent from that summary rather than the full transcript. Anthropic's guidance on tuning a compaction prompt is specific about order: get recall right first — make sure nothing important is dropped — then tighten for precision once you're confident nothing's missing. Clearing stale tool results, by contrast, is described as one of the lightest-touch, safest forms of compaction available, which is a reasonable place to start before reaching for a full summarization step.

Sub-agents keep their own mess to themselves

This is the technique Anthropic points to for parallel exploration: several sub-agents can each burn through their own scratch work without ever polluting the main agent's window with somebody else's dead ends. Anthropic's write-up cites a demo agent playing Pokémon and maintaining task precision across thousands of game steps using this combination of techniques.

Structured notes, kept outside the conversation

The last piece is low-tech on purpose: an agent keeping a running external note — a to-do list, a NOTES.md file — instead of relying on the conversation itself as its only memory of what it has and hasn't done. Anthropic's own Claude Code is the example cited for this pattern. It costs nothing to build and it doesn't compete with the context window for space, because it isn't in it.

Check the date. Context editing and the memory tool are documented as a public beta as of this writing (Anthropic dated its own posts September 29, 2025, describing Claude Sonnet 4.5). Beta features change quickly — read the linked pages directly for current availability rather than relying on this summary months later.