The post describes five techniques: offloading large tool results to files, summarizing older conversation messages, loading skills progressively, isolating noisy work in subagents, and persisting long-term memory in files. It gives implementation details, including offloading tool results above roughly 20,000 tokens and triggering summarization at 85% context-window usage by default. LangChain says memory is stored as plain files in storage controlled by the operator, allowing the same memory to be used across models or providers. The post presents a vendor-authored implementation guide, not an independent evaluation of effectiveness.
Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context
In one minute
LangChain published a technical guide to context management in its open-source Deep Agents harness.