Anthropic connects Claude’s memory across chat and Cowork—and makes control part of the product

Key points

• Claude now shares memory between chat and Cowork so users do not have to repeatedly brief the agent when moving from research to action. Sources: N1
• Users can inspect, edit, or delete retained topics, while several sensitive categories are excluded by default. Sources: N1

Continuity makes an agent feel like one system

The practical value of an agent depends on how much context it can carry into the next task. Shared memory reduces repeated setup and makes chat and Cowork behave more like one continuous assistant.

That convenience also creates a clearer governance surface. A user needs to know what was retained, correct it when it is wrong, and remove it when the context should not persist.

Sources: N1

Safety is moving into everyday product controls

Anthropic’s earlier agent and cybersecurity releases emphasized stronger capabilities under tighter controls. The memory update brings the same question into ordinary use: whether safeguards are understandable at the moment a system remembers and acts.

The competitive test is no longer only which agent completes the longest task. It is whether the product can preserve useful context without making privacy and permission boundaries invisible.

Sources: N1, S1, S2

Why it matters

Persistent context can materially improve agent adoption, but memory mistakes can also compound across tasks. Anthropic is treating user visibility and control as part of the feature; the evidence to watch is how reliably those controls work at scale.

Sources: N1

Sources

1. Claude Cowork finally remembers what you told the app in chat — TechCrunch · 2026-08-25
2. Claude Fable 5 and Claude Mythos 5 — Anthropic · 2026-06-09
3. Claude announcements: production agents and platform updates — Anthropic · 2026-08-20
4. Anthropic IPO filing will show AI backlash as a risk, sources say — CNBC · 2026-08-21