Tech • AI • Robotics
Claude Has Changed, We Need to Catch Up
8/10TL;DR
Anthropic’s Claude ecosystem has been reshaped by a new memory system, evolving model roles, a built-in browser, stronger Gmail actions and upgraded voice features, changing how users need to work with the platform.
KEY POINTS
Memory becomes a control layer
A new memory setting is being rolled out across the Claude app, with options to generate memories from conversations, reference past chats and import memory from other AI services. This is not limited to simple profile facts. It increasingly captures working habits, preferences and behavioral cues, then uses them to shape future outputs.
Sensitive data can influence responses
Users can allow or block storage of sensitive topics such as health, politics, religion or intimate personal details. That choice matters because remembered sensitivities can affect how Claude frames answers and decisions later, including in professional contexts. The feature can therefore improve personalization while raising clear privacy and workplace-governance concerns.
Past conversations are becoming operational data
The new memory system is designed to pull weak signals from earlier exchanges, such as preferred writing style, tolerance for long sessions or attitudes toward clients and tasks. In chat mode, this can let Claude infer patterns from prior discussions and adapt automatically. The feature appears unevenly deployed, with conversation search working more reliably in chat than in project-based workflows.
Model roles have shifted
The summer updates changed the practical division of labor between models. Sonnet 5 now appears to fill much of the role previously associated with Opus 4.8 for serious reasoning work, while Opus 5 is better understood as a more autonomous agent built to execute tasks, not merely assist step by step. That distinction helps explain why many users perceive it as unpredictable.
Why Opus 5 feels “rebellious”
Opus 5 tends to pursue the objective it infers, even if that means changing documents or taking initiative beyond the literal prompt. In accounting, for example, a request to create one invoice could lead it to revise numbering across older invoices if it concludes the broader system is wrong. This is less a rise in hallucination than a shift toward autonomous problem-solving.
Memory may work better than heavy prompting
For Opus 5, detailed prompts alone may not be the most effective steering method. Storing durable rules in memory, such as never altering accounting records without approval, appears to constrain behavior more naturally and reliably. That makes memory a practical governance tool, not just a convenience feature.
Claude now has two browser modes
Claude can still control the user’s own browser, opening tabs in Chrome or another installed app. That method is slower but benefits from existing cookies, sessions and saved logins. A newer internal browser is faster for form-filling and repetitive web tasks, but it does not automatically inherit the user’s authenticated sessions.
Security and speed trade off in web automation
The external-browser mode avoids handing passwords directly to Claude because it reuses the user’s logged-in environment. The internal browser is more efficient for actions such as posting listings or completing online forms, but users may need to log in separately site by site. This creates a clear trade-off between operational speed and credential handling.
Gmail actions are now far more powerful
The Gmail connector has expanded beyond reading mail, sorting messages and drafting replies. It can now write and send emails, with confirmation steps, making fully automated workflows more realistic. Combined with browser control and file access, this opens the door to end-to-end administrative tasks such as listing products, monitoring inquiries and replying to prospects.
Mobile voice features are also advancing
The mobile app is gaining stronger voice capabilities beyond simple dictation, although deployment remains uneven. Alongside browser and connector upgrades, that points to a broader shift: Claude is moving from a text assistant toward a multi-interface work agent embedded across desktop, mobile and connected tools.
CONCLUSION
The most significant change is not a single new model but a new operating logic for Claude. Users who understand memory, autonomy and tool access are likely to get far more value from the platform than those treating it as an older-style chatbot.
Explain this