Anthropic adds context editing and a file-based memory tool to Claude Developer Platform
Anthropic introduced two context management capabilities for the Claude Developer Platform alongside Claude Sonnet 4.5: context editing, which prunes stale tool calls and results as token limits approach, and a client-side, file-based memory tool that lets agents persist information outside the immediate context. These additions aim to help agents sustain long-running workflows without losing critical information or hitting context limits.
How context editing works
Context editing automatically removes outdated tool outputs and other stale content from the conversation when token usage approaches the model’s limits, while preserving conversational continuity. The mechanism reduces clutter in the active context so the model focuses on recent, relevant material. Claude Sonnet 4.5 adds built-in context awareness that tracks available tokens throughout a conversation to make pruning decisions more effective.
The memory tool: persistent, client-side storage
The memory tool exposes a file-based interface for agents to create, read, update, and delete files in a dedicated memory directory. Storage and persistence are managed by the developer’s infrastructure, keeping data and retention policies under direct control. Because the tool operates entirely through tool calls, agents can offload larger or longer-lived knowledge than the context window permits, then consult that stored knowledge across sessions.
Measured effects on agent workflows
Internal evaluations reported meaningful gains when combining the two features:
- 39% performance improvement when using both the memory tool and context editing versus baseline.
- 29% improvement with context editing alone.
- In a 100-turn web search test, context editing enabled completion of workflows that would otherwise fail from context exhaustion and reduced token consumption by 84%.
Practical use cases
These capabilities are oriented toward agentic tasks that span extended interactions:
- Coding: pruning old file reads and test outputs while persisting debugging notes and architectural decisions to memory for later reference.
- Research: preserving key findings in memory while expiring outdated search results from active context.
- Data processing: storing intermediate results persistently and clearing bulky raw data from context to keep pipelines within token budgets.
Availability and resources
The context management features are available in public beta on the Claude Developer Platform, and are also supported in Amazon Bedrock and Google Cloud’s Vertex AI. Documentation and examples are available for developers:
- Context editing: https://docs.claude.com/en/docs/build-with-claude/context-editing
- Memory tool: https://docs.claude.com/en/docs/agents-and-tools/tool-use/memory-tool
- Cookbook examples: https://github.com/anthropics/claude-cookbooks/blob/main/tool_use/memory_cookbook.ipynb
Original source: https://www.anthropic.com/news/context-management
