Context engineering: memory, compaction, and tool clearing Mar 2026 • Tools Agent Patterns Compare context engineering strategies for long-running agents and learn when each applies, what it costs, and how they compose.

This cookbook teaches context engineering strategies for long-running AI agents, focusing on three key techniques: compaction (summarizing context), tool-result clearing (removing re-fetchable tool outputs), and memory (persistent external storage). The guide addresses context rot—the degradation of model performance as context windows grow—and provides practical implementations using Claude's API. Through a research agent example, it demonstrates how to combine these strategies to manage token growth, maintain conversation continuity, and persist knowledge across sessions.

Key Points

• •Context rot occurs as token count increases, reducing model's ability to recall information accurately before hitting hard token limits
• •Compaction distills context into high-fidelity summaries, allowing agents to continue with minimal performance degradation on long conversations
• •Tool-result clearing removes old, re-fetchable tool outputs (file reads, API responses) while maintaining call records to reduce context bloat
• •Memory implements structured note-taking via persistent external storage, enabling agents to track progress across tasks and sessions without keeping everything in active context
• •All three strategies have first-party API support: server-side compaction, context editing (tool-result clearing), and the memory tool
• •Map workload characteristics to the right primitive: use clearing for large re-fetchable results, compaction for long conversations, memory for cross-session persistence
• •Claude Code uses multiple strategies in production: compaction for conversations and dual memory systems for cross-session persistence
• •Context is a finite resource with diminishing marginal returns; the goal is finding the smallest set of high-signal tokens maximizing desired outcomes
• •Subagents isolate work in separate contexts and programmatic tool calling keeps large results out of the window entirely as complementary strategies
• •Test clearing configs and compaction prompts against your workload's actual tool-use patterns to diagnose which context problem needs solving

Workflow Diagram

Concepts

Artifacts (4)

ANTHROPIC_API_KEY=your-key-here

pip install anthropic python-dotenv matplotlib

import json
import os
import tempfile
from collections import namedtuple
from pathlib import Path
import anthropic
import matplotlib.pyplot as plt
from dotenv import load_dotenv

load_dotenv()

if not os.environ.get("ANTHROPIC_API_KEY"):
    raise ValueError("ANTHROPIC_API_KEY not set. Add it to a .env file or export it.")

CORPUS_PATH = Path("research_corpus.py")
assert CORPUS_PATH.exists(), (
    f"research_corpus.py not found in {Path.cwd()}. It should be alongside this notebook."
)

client = anthropic.Anthropic()
MODEL = "claude-sonnet-4-6"

print(f"anthropic SDK {anthropic.__version__}, model {MODEL}")

# Token counting utility
_token_cache: dict[str, int] = {}

def count_tokens(text: str) -> int:
    if text not in _token_cache:
        _token_cache[text] = client.messages.count_tokens(
            model=MODEL,
            messages=[{"role": "user", "content": text}]
        ).input_tokens
    return _token_cache[text]

from research_corpus import COMPACTION_PROBES, CORPUS

print(f"CORPUS is a dict of {len(CORPUS)} synthetic documents held in Python memory.")
print("When the agent calls read_file, the content is served from this dict and")
print("lands directly in the agent's context window — no disk I/O involved.\n")

_total_tokens = 0
for path, content in CORPUS.items():
    n_tok = count_tokens(content)
    _total_tokens += n_tok
    display_name = path.removeprefix("/research/")
    print(f"{display_name:<26} ~ {n_tok:>6,} tokens")

print(f"\nTotal corpus: ~ {_total_tokens:,} tokens")

assert _total_tokens > 250_000, (
    f"Corpus is only {_total_tokens:,} tokens; expected >250K. "
    "Restart the kernel and re-run, or verify research_corpus.py is current."
)