Anthropic Launches AI Memory Tool and Context Editing in Public Beta
Anthropic, the AI company behind Claude, has launched two significant features in public beta: a memory tool and context editing. These are now available on the Claude Developer Platform, as well as through Amazon Bedrock and Google Cloud Vertex AI, with Anthropic providing detailed documentations and a cookbook.
The introduction of these tools marks a shift in AI strategy, focusing on 'just in time' data processing and minimal overlaps in functionality. This approach aims to maximize token efficiency, a crucial aspect in AI tool development.
Anthropic has also revisited and popularized the term 'context engineering', first used by Riley Goodside in early 2023. This term, later championed by Shopify CEO Tobi Lütke and former OpenAI researcher Andrej Karpathy in 2025, encompasses managing the entire set of tokens used during model inference. This includes system instructions, tools, external data, and message history, providing a more comprehensive approach than traditional prompt engineering.
For system prompts, Anthropic advises finding the right 'height level' - specific enough for behavior control, but flexible enough for strong heuristics. The combination of the memory tool and context editing has significantly improved agent search performance by 39 percent, with context editing alone bringing a 29 percent improvement. Moreover, it has reduced token consumption by 84 percent in a 100-round web search.
Anthropic has developed three main techniques for multi-step tasks: compaction, structured notes, and sub-agent architectures. These techniques help AI agents efficiently use their limited 'attention budget', preventing context decay that occurs when processing large context amounts.
The public beta launch of the memory tool and context editing features, along with the renewed focus on context engineering, signals a significant advancement in AI capabilities. These tools and techniques enable AI agents to maintain coherence in long-term tasks and use their limited attention more efficiently. As AI continues to evolve, such innovations are crucial for pushing the boundaries of what AI can achieve.
