Artificial Intelligence (AI) assistants are shifting from isolated chat windows to persistent companions with memory and access to other applications. This affects you when you want more than answers and begin delegating tasks such as organizing appointments, preparing documents, or managing longer workflows. New features from Anthropic and OpenAI show why convenience, privacy, administration, and cost can no longer be assessed separately.
From chat to persistent assistance
An AI agent is a system that plans a task and carries out multiple steps instead of merely answering one question. A persistent agent also retains information from conversations or activities and can reuse it later. You no longer have to explain projects, preferences, and organizational details whenever you move from research to execution.
Anthropic is addressing this by merging memory across Claude Chat and Claude Cowork. According to the report on Claude’s shared memory, Cowork can use information from earlier chats while completing practical tasks. Memories are updated by topic during an active conversation rather than summarized only after that conversation ends.
One example is preparing a conference agenda. If you previously discussed the headcount, location, and speakers, Cowork should be able to account for those details when drafting the agenda. The agent could likewise prepare an update for your manager without requiring you to enter the entire project background again. This can save time in recurring office work and make chatting and task execution feel like one continuous service.
OpenAI is pursuing a similar transition from answering to acting. ChatGPT Work is designed to handle multistep projects for professions beyond software development and, according to TechCrunch’s report on OpenAI’s agent strategy, is available on the lowest subscription tier mentioned at $20 per month. Its value, however, depends on giving the system access to work tools such as email, messages, notes, or design applications.
Why memory is useful and sensitive
Shared memory reduces repetition and may improve results because an agent knows earlier decisions and terminology. Anthropic says it also exposes the topics Claude has retained, allowing you to read, edit, or delete them. That is a meaningful control because a memory system that remains invisible would be difficult to audit.
Anthropic also says Claude will not store personal or sensitive details such as health information, religion, political views, or gender identity by default. If you deliberately enable sensitive topics, the application is supposed to notify you when one is saved. Certain categories, including government-issued identification, Social Security numbers, criminal history, and immigration status, are not supposed to be retained as memories at all. These are provider claims and have not been independently verified in the supplied sources.
OpenAI’s Computer History goes beyond conversational memory. According to a hands-on test of Computer History, the optional feature in the official ChatGPT application for macOS records computer activity. The report says it does not store screenshots, although it apparently reads them, and allows you to exclude individual applications and websites.
Computer History is available to individuals through the Pro plan for about 100 euros and to Business and Enterprise customers when an administrator enables it. Whether collected data is used for model training apparently depends on the account’s general training setting, according to the test. The report describes protection of the data on the Mac as weak. This does not directly contradict Anthropic’s approach, but it shows that “memory” can mean very different things: selected topics in one product and a broad activity history in another.
Pros and Cons of persistent AI agents
Pros:
- Less repetition – You do not have to reenter project details and earlier decisions for every task.
- Continuous workflows – Research, planning, and execution can remain connected across chat and work modes.
- Practical assistance – Agents can book appointments and reservations, arrange an airport ride, or organize an inbox.
- Manageable memories – Visible entries that you can edit or delete provide at least one direct form of control.
Cons:
- Extensive visibility – Email, messages, screens, location, and calendars can combine into a detailed picture of your daily life.
- Unclear context boundaries – An agent might include information from a private direct message in a work document if it misjudges confidentiality.
- Binding actions – Depending on the service, an agent may enter agreements or transactions on your behalf.
- Uncertain costs – Multistep tasks require substantially more model activity than a short chat.
Instinct, an assistant still in private testing, illustrates how broad this access can become. According to a report on Instinct’s privacy model, it connects to email, messaging, calendars, audio, location, screens, and other sources. Testers used it for tasks including booking appointments and reservations, organizing shopping, and finding inexpensive flights.
The terms cited in the source also grant the operator a broad, perpetual, and irrevocable license to user materials, including for training models. They allow Instinct to enter binding agreements, commitments, or transactions on a user’s behalf. Because the service remains in private testing, these observations do not establish how it would behave at wider scale. They do demonstrate why capability and access policies need to be evaluated together.
How you can keep control
If you are new to agents, begin with a narrowly defined task that you can reverse. Drafting a conference agenda from known details is a more sensible first test than allowing an agent to book travel and accommodation independently. Review both the result and any saved memories before connecting further applications.
Advanced users can get more value by separating tasks, data spaces, and permissions. An agent used for general research does not automatically need private messages, location data, or your entire inbox. In a business setting, you should also define which steps may run automatically and which require approval from a person.
Step 1: Limit the task and its data
- Choose a recurring workflow whose result you can easily verify.
- List the information it actually needs and the details that must remain confidential.
- Connect only the necessary application at first, or provide the required details manually.
Step 2: Review memory and permissions
- Read stored topics and remove information that is inaccurate or unnecessary.
- Check the training setting, exclusion lists for applications and websites, and options for sensitive memories.
- Allow bookings, messages, or other external actions only where you can detect and correct mistakes.
Step 3: Administer usage
- For teams, assign responsibility for members, permissions, and usage limits.
- Regularly review which workspaces and data sources an agent actually uses.
- Compare the time saved with subscription charges, usage costs, and the effort required for oversight.
For organizations, this oversight becomes a task of its own. OpenAI’s Admin plugin for ChatGPT Work and Codex is intended to analyze workspace usage, manage members and permissions, adjust limits, and act on administrative requests. The supplied summary provides no details about logging, privacy, or pricing, so the quality of those controls cannot yet be assessed comprehensively.
What costs and administration change
Agents create more computational work because they plan, execute intermediate steps, check results, and repeatedly call an AI model. A token is a small unit of text used to measure model processing. According to data on token use by agents, agentic applications on OpenRouter recently consumed almost five times as many tokens as direct human use.
The seven-day average on that platform was 7.3 trillion tokens, while agentic usage had reportedly risen roughly fourteenfold since February, based on figures attributed to Andreessen Horowitz. These numbers relate to the specified platform and cannot automatically be generalized to the entire market. They nevertheless show how an agent can create substantial background activity despite requiring only a few visible instructions from you.
Gartner expects the cost per agentic workflow to rise to more than five times its current level by 2028, according to a Netzwoche report on rising inference costs. Inference is the computational process through which a trained model produces a new response. Assigning a task to an agentic reasoning model already creates at least five times the provider-side inference cost of a simple chatbot interaction, according to Gartner.
This may sound contradictory because individual tokens are becoming cheaper. Gartner attributes the effect to more capable models and more complex workflows that consume far more tokens overall. The estimate is a forecast, not a confirmed price trajectory. For you or your organization, it means that a monthly subscription price is not a sufficient cost measure: limits, task scope, and the number of automated steps also belong in the calculation.
The sources provide no separate Swiss prices, language restrictions, or firm availability details. Computer History became available in the European Union after an initially limited rollout, but that does not establish the same rollout status in Switzerland. Swiss individuals, schools, and organizations therefore need to check the availability, storage, training, and administration options shown in their own accounts without assuming that an EU release determines local privacy implications.
Persistent AI agents are most useful when you handle recurring tasks with stable context and can verify their output. That same context makes them more data-intensive, while extensive permissions increase the impact of mistakes or unclear terms. The unresolved issue is whether providers can improve transparent control and predictable costs as quickly as they expand memory and autonomy.
Sources
- Claude Cowork finally remembers what you told the app in chat – TechCrunch, 2026-08-25
- Computer History: ChatGPT speichert auf Wunsch alles, was du am Mac tust – wir haben es ausprobiert – t3n, 2026-08-24
- Introducing the Admin plugin for ChatGPT Work and Codex – OpenAI, 2026-08-25
- Instinct’s powerful AI assistant is raising privacy and security concerns – TechCrunch, 2026-08-24
- OpenAI is building AI agents for everything. Will everyone use them? – TechCrunch, 2026-08-24
- KI-Agenten verbrauchen 5-mal mehr Token als Menschen – das ist der Grund – t3n, 2026-08-24
- Inferenzkosten für KI-Agenten steigen bis 2028 auf mehr als das Fünffache – Netzwoche, 2026-08-24
