Three API Primitives Cut a Research Agent's Context From 335K Tokens to 50K. Here's How.

Context engineering just got its cookbook.

Anthropic's Claude Cookbook published a complete guide to three API-level primitives that, used together, reduced a research agent's peak context from 335K tokens to a sustained 50-80K range. Server-side compaction. Tool-result clearing. Persistent memory. Each one is simple. Together they change the economics of long-running agents.

Compaction triggers at a configurable token threshold (the cookbook uses 150K) and distills the conversation to a summary, preserving key decisions while dropping the narrative. Tool-result clearing replaces bulky tool outputs with [cleared] placeholders at zero inference cost, because tool results are almost always re-fetchable. Persistent memory stores cross-session state in files the model manages itself.

The diagnostic framework is the part I keep thinking about. If more than 80% of your context is re-fetchable tool output, use clearing first. If long dialogue accumulates, add compaction. If work spans sessions, add memory. That's not a recipe. It's a decision tree. And it works because it forces you to ask the right question: what in this context window is actually load-bearing?

I run a research agent pipeline every day in my personal projects. Context bloat is the single biggest reliability problem I hit. Agents get dumber as context fills up. They start ignoring earlier instructions. They lose track of what they've already done. The fix isn't a bigger context window. It's being surgical about what stays in it.

The cookbook also introduces the "Summarize up to here" option in Claude Code's Rewind menu. Manual mid-conversation compaction. Use it after completing a major milestone to reclaim token budget while preserving key decisions. Claude Code docs confirm this shipped recently.

For builders: implement tool clearing before anything else. If your agents call tools that return large JSON payloads, you're paying to re-read that data on every subsequent turn. Clear it after extraction. The cost savings compound with every interaction.