Claude's AI Agents Can Now Remember Between Sessions

Anthropic just made its AI agents considerably more useful, and the change is worth serious attention from any UK business that relies on repetitive or knowledge-heavy processes.

Memory on Claude Managed Agents entered public beta on 23/04/2026. It gives Claude agents a persistent, file-based memory layer that survives the end of each session. Instead of starting from a blank slate every time, an agent now carries forward what it has learned: client preferences, workflow patterns, document quirks, and operational history.

Until this update, every Claude Managed Agent session was stateless. Developers who wanted continuity between sessions had to build their own memory scaffolding, which meant custom databases, retrieval logic, and a significant amount of extra engineering work. For teams without a large in-house AI function, that overhead often killed projects before they got started.

The new memory layer removes that burden. Memories are stored as files on a mounted filesystem, auditable and exportable via the API. Every session carries a full audit trail, and organisations can roll back specific memories, redact data, or delete records outright. That level of control matters enormously in regulated sectors, where demonstrating what an AI system knew, and when, is a compliance requirement rather than an optional extra.

Enabling memory requires no separate access request. It is available under the standard managed-agents-2026-04-01 beta header for anyone already using the Claude platform.

The practical benefits show up quickly in document-heavy and repetitive workflows. Wisedocs, one of Anthropic's early adopters, reported a 30 percent speed increase in document verification after deploying agents with persistent memory. Rakuten reported a 97 percent reduction in first-pass errors. These are not marginal improvements; they represent the difference between an AI tool that helps at the margins and one that materially changes how work gets done.

For UK professional services firms, finance teams, and operations departments, those numbers carry real weight. Fewer errors in document processing means fewer write-offs, fewer compliance incidents, and less time spent on manual review. Faster turnaround means better client experience without adding headcount.

There is also a longer-term benefit worth naming directly. An agent that learns becomes more useful as time goes on. A support agent that remembers returning clients does not repeat the same onboarding questions every session. A finance agent that has processed hundreds of invoices for a particular client learns its patterns and flags anomalies faster. The longer the agent runs, the more value it returns, without additional development effort.

Because all memory changes are fully logged, businesses operating in regulated environments can demonstrate to auditors exactly what an agent knew, when it knew it, and whether that knowledge has since been retained, amended, or removed. That is a considerably stronger compliance position than most AI workflows currently offer.

At Adevious AI we have been building on the Claude platform since its early access period. The Managed Agents framework has already improved how we think about agentic task design, but stateless sessions have been the most consistent friction point we encounter with clients. Building memory scaffolding from scratch is genuinely time-consuming, and it introduces dependencies that need ongoing maintenance. This release closes that gap for the majority of use cases. For small and mid-sized UK businesses that want to deploy reliable, self-improving AI agents without a large internal engineering team, the infrastructure is now managed for you: secure sandboxing, long-running sessions, persistent memory, and audit logging, all in one place and priced against actual usage.

The cost is $0.08 per session-hour on top of standard Claude API token rates. For a business running agents a few hours per day, that overhead is modest against the workflow gains on offer.

If your team already uses the Claude API, read the memory documentation at platform.claude.com and enable it on an existing Managed Agent configuration. Run it against a workflow that currently involves repetitive manual context-setting and note the difference after a week of operation.

If you are not yet using Claude agents but are thinking about whether AI automation could reduce bottlenecks in your business, this is a good moment to explore a pilot. The complexity of getting started has come down considerably over the past few months. If you would like to talk through where persistent AI agents could add the most practical value for your team, get in touch with Adevious AI. We work with UK businesses of all sizes to identify the right starting point and build from there.