Today Anthropic introduces two tools to help agents built on Claude handle information better: context editing and the memory tool, combined with the new model Claude Sonnet 4.5.

These features aim to let agents run longer without losing important details or clogging the context window. Curious how that works in practice? Anthropic explains it here. (anthropic.com)

What changes for developers

The big struggle for agents today isn’t creativity — it’s working memory. When an agent runs many tasks and produces tool outputs, the context window fills up and you have to decide what to keep and what to discard.

With context editing, Anthropic automates that cleanup: it removes tool calls and results that are already stale as the token limit approaches, while keeping the conversational flow intact. That extends an agent’s useful life without you having to intervene manually. (anthropic.com)

The memory tool complements this approach. Instead of trying to hold everything in the context window, Claude can create, read, update, and delete files in a memory directory that persists outside the context. That storage is accessed via tool calls and lives in your infrastructure, so you control where and how data is kept. In short: less noise in the conversation and more useful info stored long-term. (anthropic.com)

Concrete use cases

• Coding: imagine an agent scanning a large repo. Context editing drops old file reads and test results, while memory stores design decisions and debugging clues. The result: the agent doesn’t “forget” because the context got full. (anthropic.com)

• Research: the agent can accumulate key findings in memory and clean up old searches, building a knowledge base that improves over time. (anthropic.com)

• Data processing: save intermediate results to memory and clear raw data from the context so long pipelines complete without token exhaustion. (anthropic.com)

Results and metrics

Anthropic reports performance gains in their internal tests. In an evaluation set for retrieval-based agents, combining the memory tool with context editing improved performance by 39% over the baseline. Context editing alone gave a 29% improvement.

In a 100-turn test, context cleanup allowed flows that would have failed due to exhaustion to complete, and reduced token usage by 84%. These figures come from Anthropic’s published internal evaluation. (anthropic.com)

How to get started

These capabilities are available in public beta on the Claude Developer Platform and also via Amazon Bedrock and Google Cloud Vertex AI. If you want to try them, Anthropic publishes specific documentation on context editing and the memory tool and offers examples in their cookbook.

In practice you’ll need to decide on the storage backend and retention policies, because memory lives on the client side. That gives you control — and responsibility — over the data. (anthropic.com)

Preserving what matters and deleting what’s extra is today the most practical way to scale intelligent agents.

Final reflection

And what if you’re not a developer? This isn’t only for engineers. These improvements mean assistants that remember long projects, support tools that don’t lose track of an incident, and apps that handle files and data without restarting the conversation every few minutes.

Of course, privacy and governance matter: memory is stored in your infrastructure, which gives control but also requires careful policies.

If you have a project where an agent runs out of context or repeats work, these tools change the rules. It’s not magic — it’s practical context management.