Instinct AI: how it works, what it remembers, and what a work version needs

Instinct AI is a personal assistant you can text or call to handle everyday tasks, from ordering groceries to booking travel, according to its official site. It can work with information you share and accounts you connect, but using it for company work requires checking how memory and permissions behave.

The attraction is easy to understand: hand over the errand and keep using your existing messages. For a founder, that might mean less time arranging travel between meetings. Putting customer records into the same assistant introduces a different set of decisions about who controls the information and what survives a handoff.

We checked the public product pages and policies for this article. We also read published hands-on reporting and an outside memory teardown; we haven't run our own account-level evaluation of Instinct. Where the internal design remains uncertain, we keep that uncertainty attached to the claim.

What does Instinct AI do?

Instinct describes a personal assistant that understands what you're working on and uses a phone and computer to complete tasks. Its official homepage gives concrete examples: ordering recipe ingredients, booking a train and sending flowers. Those are useful jobs when the alternative is navigating several websites yourself.

The Verge's hands-on report describes conversations through iMessage, WhatsApp and email. The reviewer used it for appointment requests and return paperwork. That reporting supports the personal-assistant framing, while leaving room for errors in the details of a task.

The same review found that Instinct missed a prerequisite when researching swimming lessons. The suggested sessions looked plausible, but a condition on the provider's website changed what the user could book. An assistant can finish the visible steps while overlooking a rule that determines whether the result is usable.

For work, start with tasks whose results you can inspect. Researching travel options or preparing a draft gives you a chance to catch a mistaken assumption before it becomes a commitment. Treat permission to make a purchase or contact a customer as a separate decision.

What does Instinct remember?

Instinct's policies describe using information from your interactions and connected services to personalize the assistant. The most detailed public account of its memory architecture comes from an outside investigation, rather than official technical documentation. That distinction matters when deciding what to trust about persistence or deletion.

Supermemory's memory teardown inferred linked Markdown files tracked with Git, alongside summaries used to give the answering model context. Markdown is plain text with lightweight formatting; Git records changes to files. The proposed design could explain how an assistant retrieves a preference mentioned in an earlier conversation.

The author explicitly says the reconstruction came from probing the assistant and that they hadn't seen its code. Supermemory also sells memory infrastructure, so its interpretation comes from a vendor with a commercial interest. We wouldn't turn the teardown into a claim that Instinct's current implementation follows a particular storage or retrieval design.

Public evidence	What it supports	What still needs checking
Instinct's privacy policy describes connected information and personalization	The service processes context to help perform tasks	Which specific facts remain available in your account
Supermemory reports finding linked records and summaries	A plausible explanation of persistent memory	Whether that reconstruction matches the current implementation
Hands-on reporting describes useful contextual recall	The assistant can bring earlier information into a task	Reliability across your own recurring work
The privacy policy describes disconnection and deletion controls	There are documented ways to manage connected information	What each control removes and what remains afterward

For an operator, recall needs a source and a way to correct it. A useful assistant should distinguish an old customer request from an agreed contract term. Remembering both without identifying which governs the next action can produce a confident, expensive mistake.

How should you test its memory for work?

First confirm that your intended use and evaluation are permitted. Instinct's terms of service grant personal use only, although their introduction also contemplates company representatives. They restrict benchmarking without written permission, subject to applicable law, so clarify business use and any proposed evaluation with Instinct before proceeding.

For an assistant you're authorized to evaluate, use a harmless, fictional work example before connecting sensitive records. Imagine a renewal where a customer requested a discount and your company later declined it. Ask the assistant to explain the current position and identify the evidence behind its answer.

Then supply a correction and repeat the task in a later conversation. Check whether the correction changes the recommendation and whether the assistant can still distinguish the earlier request from the final decision. You're evaluating a work outcome, so a convincing description of its memory system doesn't settle the test.

Also test missing information. Give the assistant an incomplete meeting recap and ask what can safely go into a follow-up email. It should preserve uncertainty around anything the recap doesn't establish, particularly promises about scope or delivery.

Keep a written record of the expected answer and what happened. When the assistant gets something wrong, record whether you could fix the underlying information or merely persuade it to answer differently in that conversation. That difference becomes important when repeat work runs while you're away.

What happens to your data?

Instinct processes information you provide and information made available through connected accounts, subject to the permissions you grant. Its privacy policy says that can include private communications and sensitive information. An account connection therefore deserves the same review you'd give another service that can read company records.

The policy allows certain information to be used to develop and train models. It offers an opt-out for future model training, with an exception for information flagged for safety review. It also says the opt-out doesn't reverse prior training and doesn't cover every other form of service improvement.

The same policy separately says information received from Google Workspace APIs isn't used to train or improve AI models. Preserve that distinction when evaluating the product. A broad statement that all connected data trains models would misrepresent the published policy.

Instinct's terms also provide a Vault protection: materials you add there are used only to provide the service and aren't used to train AI models. That contractual promise is distinct from the general training opt-out. Confirm which materials enter the Vault before assuming it covers everything you share.

Disconnection and deletion also have different effects. Instinct says disconnecting an integration doesn't automatically delete information already collected, and it describes controls for removing indexed external information. Before adding company data, establish which action stops future access and which removes existing records.

These are published terms, rather than an independent security assessment. A company should still confirm the contractual treatment of its information and decide which records employees may share. Keep sensitive material out of a trial until whoever owns that decision has reviewed the terms.

What should an assistant ask before doing?

Separate permission to read information from permission to act on it. An assistant may need a receipt to investigate a return, but that doesn't mean it should accept a replacement product or an additional charge. Write down the boundary before starting the task, especially when the assistant will interact with another business.

For a hypothetical customer follow-up, you might allow the assistant to gather the relevant emails and prepare a reply. Keep sending behind your approval until you've checked the recipient and the commitments in the draft. A correct summary can still produce an unsuitable message if it assumes a concession the company hasn't authorized.

Instinct's privacy policy acknowledges the possibility of unintended payments or communications, as well as misleading instructions from third parties. That is a reason to evaluate the controls available in your account and watch early tasks closely. Reading a webpage shouldn't give that webpage authority to change the job you assigned.

Test a refusal boundary alongside the successful task. If you authorize research but withhold permission to buy, check that the assistant stops at a recommendation even when a booking page urges immediate payment. If the current controls can't enforce the boundary you need, reduce the permissions or keep that work manual.

Finally, decide how you'll recognize completion. For a booking, the evidence could be a confirmation from the provider; for a draft, it could be text you review against the original thread. Ask the assistant to report a blocker plainly when it can't finish, and check that report rather than assuming silence means success.

Can Instinct work with colleagues?

Published reporting describes Instinct assistants coordinating for trusted friends and coworkers. SiliconANGLE's coverage includes sharing files and organizing schedules between assistants. It would be unfair to describe Instinct as limited to isolated conversations with a single user.

Coordination still leaves questions about company ownership. An employee can authorize their assistant to share a file, while the company may need a lasting record that another employee can maintain. Evaluate whether the current product supports the particular handoff your company requires.

For example, consider an account manager leaving the business. The replacement needs the agreed renewal terms and their sources, while the departing employee's personal appointments should stay private. Ask the vendor to demonstrate how that separation works before treating a personal assistant as a company-wide source of truth.

Our guide to what a company brain contains explains the shared-record approach. The practical requirement is that useful context remains available to the right people when responsibilities change. A text-message interface can be convenient without answering that ownership question by itself.

What does Instinct cost?

Instinct's public pages didn't give us a published subscription amount when checked for this article. Its pricing URL showed no plan details, while its terms of service allow charges for paid services. We can't turn that into a current, guaranteed free plan.

The Verge's October hands-on report quoted Instinct saying the service was free for invite-only users. That is useful dated evidence about early access. Confirm the offer in your account, including who pays for purchases the assistant makes, before relying on it in an operating budget.

A free trial can establish whether you like handing off errands. It doesn't establish what repeat business use will cost later or what contractual controls are included. Avoid making a long-term workflow dependent on an assumed price that the vendor hasn't published.

What does a work version need?

A work assistant needs access to a maintained company record, with permissions that match people's responsibilities. It should let an operator check the source behind an answer and control consequential actions. Those requirements apply whether the interface is a text conversation or a web app.

Hockeystick takes that shared-record approach: our brain holds linked notes about company activity, and Ask returns answers with sources. Our current Gmail and Google Calendar connection supplies read-only context for filing. It doesn't itself send email or edit calendar events, so those actions need a separately configured connection and approval.

On the live site, Hockeystick's agent asks before sending, posting, deleting or spending; scheduled jobs can't do those things. Each customer has their own agent and brain app in a private space. Those are specific controls to evaluate against your work, rather than a promise that AI can't make mistakes.

According to Hockeystick's current pricing, Personal is $49 a month with $10 of included AI, while Team is $199 a month flat for unlimited people with $100 of included AI. Team members share the brain and can ask it questions or discuss records in threads. The agent currently works for the owner; buying Team doesn't give each employee their own assistant.

Our applications directory also distinguishes a listed app from a tested connection. A listing marked "Request it" shouldn't become an assumption that the app works today. Bring the exact read or write action your workflow requires into the setup discussion.

Who should choose Instinct?

Choose Instinct when the immediate job is personal administration and its messaging interface suits how you work. Its published use cases and hands-on coverage make a reasonable case for trying it on errands you can inspect. A shared company record may add unnecessary setup if all you need is help arranging a trip.

Choose a company-focused system when the work depends on colleagues using the same maintained records. Evaluate the sources behind an answer and verify that access changes correctly when a role changes. Hockeystick is worth considering for that job, with the owner-agent limitation stated above.

For either route, start with a bounded task and check the result against the original records. Keep the workflow narrow until corrections persist and consequential actions stay within the permissions you've granted. Expand it when the actual work gives you evidence that the assistant can handle more.