Context offloading is a context engineering pattern where large or low-immediacy information is stored outside the model's active context window and replaced with a compact reference. The model can retrieve the full material later if it becomes relevant.
Typical offload targets include:
- long search result pages;
- large code files or repository listings;
- PDFs, spreadsheets, logs, and transcripts;
- generated artifacts;
- intermediate plans and scratch work; and
- repeated tool outputs.
For example, an agent might write a long API response to a file, keep only the file path and a short summary in the prompt, and read the file again only when it needs exact details. This supports context reduction without losing recoverability.
Context offloading is not the same as memory. Agent memory stores durable knowledge across tasks. Offloading usually stores working material for a specific task or run.
The main risks are stale references, missing permissions, and false confidence in summaries. Systems should preserve provenance, use stable identifiers, and make it easy for the model or application code to rehydrate exact source material.
LangChain discusses filesystem-based context management in How agents can use filesystems for context engineering.
The LLM Knowledge Base is a collection of bite-sized explanations for commonly used terms and abbreviations related to Large Language Models and Generative AI.
It's an educational resource that helps you stay up-to-date with the latest developments in AI research and its applications.