Anthropic gives Claude agents memory and context editing

If you have ever built an AI agent to tackle complex workflows, you know the dread of the context wall. After dozens of tool calls and file reads, the context window gets clogged, latency spikes, and the model starts losing the plot.

Anthropic is tackling this headache head-on. The company just introduced two new context management features on the Claude Developer Platform: context editing and the memory tool, built to give Claude Sonnet 4.5 much higher stamina on long-running tasks.

Trimming the fat while keeping the notes

Context editing basically acts as an automated garbage collector. As an agent works and approaches token limits, the system quietly strips out stale tool calls and historical outputs while keeping the conversational backbone intact.

In Anthropic's 100-turn web search benchmark, context editing slashed token consumption by 84% while completing tasks that previously crashed due to context exhaustion.

Meanwhile, the memory tool acts as an external filing cabinet outside the context window. Claude can create, read, update, and delete files inside a dedicated directory on your own infrastructure.

Instead of cramming an entire debugging history into every single prompt, Claude can jot down architectural takeaways into local files and pull them up across different sessions. On internal agentic search benchmarks, context editing alone improved performance by 29%, and pairing it with the memory tool bumped gains to 39%.

Both features are available in public beta across Anthropic's platform, Amazon Bedrock, and Google Cloud Vertex AI.

Why it matters

We think Anthropic is making the right bet by prioritizing context hygiene over just chasing gigantic context window numbers.

Brute-forcing longer context windows usually creates two nasty side effects: exploding token bills and degrading model attention. Teaching the model to actively tidy up its workspace and write notes to disk mimics how human engineers actually work through massive projects.

For developers building autonomous software engineers or deep research agents, this solves a massive architectural headache. You no longer need to build brittle custom middleware just to summarize transcripts before the context window bursts.

Keeping the memory client-side is also a clever move for enterprise compliance, as developers retain full custody over the storage backend. The real test now is seeing how well developers manage file hygiene inside that memory folder before it turns into its own digital junk drawer.