Long-Horizon Context Strategies

Long-horizon strategies handle conversations and tasks that exceed the context window. Techniques include compaction (summarizing history into a smaller representation), structured note taking (persistent external memory), and multi-agent decomposition where sub-agents handle bounded subtasks and return condensed summaries.

API entry from apis.yml

apis.yml Raw ↑
aid: context-engineering:long-horizon-strategies
name: Long-Horizon Context Strategies
description: Long-horizon strategies handle conversations and tasks that exceed the context window. Techniques
  include compaction (summarizing history into a smaller representation), structured note taking (persistent
  external memory), and multi-agent decomposition where sub-agents handle bounded subtasks and return
  condensed summaries.
humanURL: https://www.anthropic.com/news/contextual-retrieval
baseURL: https://www.anthropic.com
tags:
- Compaction
- Long Context
- Memory
- Multi-Agent
properties:
- type: Documentation
  url: https://www.anthropic.com/news/contextual-retrieval
- type: Reference
  url: https://www.anthropic.com/research/swe-bench-sonnet
- type: Reference
  url: https://github.com/microsoft/autogen
x-features:
- Conversation summarization for compaction
- Persistent memory files for cross-session knowledge
- Multi-agent decomposition with bounded sub-agents
- Hierarchical planning over long-running tasks
x-useCases:
- Long-running coding agents and SWE assistants
- Multi-day customer engagements requiring memory
- Complex research tasks decomposed across sub-agents