Wrong answers are the obvious failure. The expensive failure mode of an in-app assistant is a correct answer that forgets the customer who asked. Bybit wants Bybit AI to stop doing that.
On October 22, 2026, the cryptocurrency exchange, which calls itself the New Financial Platform, announced personalization and memory features for Bybit AI, the conversational assistant inside its app. The release, distributed by PR Newswire, casts the update as a move toward what Bybit calls AI-native financial services, with chat as plumbing for the platform rather than a help widget in the corner.
Treat it as a product announcement, not a performance study. Bybit describes intended capabilities, permissions, and limits. It does not publish accuracy figures, adoption numbers, or a design anyone outside the company can inspect.
From stateless replies to an account profile
Bybit says Bybit AI launched in September. It does not name the year. Sponsored coverage on HackerNoon reported a livestream demo on 9 September by Ben Zhou, described there as co-founder and chief executive. That date has no year either.
What the October release pins down is a shift from stateless replies to an account profile. Bybit says the assistant builds that profile from each user’s assets, positions, behavior, and risk profile, then carries the context across the app instead of treating every prompt as a cold start. The company calls the result each user’s “personal financial edge.”
Personalized. With permissions granted and limits set, Bybit AI is supposed to read the actual account (financial behavior, risk profile, assets, and open positions) and tailor what it says. The release names those inputs. It never defines what the limits block, so do not treat those boundaries as controls that have been specified.
Connected. The new claim is profile-based use of a span Bybit already advertised: one chat drawing on Markets, Trading, Earn, Card, and other products, using holdings and activity across them. Sponsored launch coverage had already described a single window across a wide set of services. This release does not show that linking products began here.
Remembered. Users can choose to share preferences. Bybit says the product retains that context through long-term memory so later conversations build on earlier ones. How long it is stored, and whether a person can view, correct, or delete it, goes unstated.
Strip the exchange away and the pattern is familiar to anyone buying AI for business. Chat becomes the front door. Memory is supposed to keep the customer. The open question is whether a path to action, and a wealth-plan prompt, can sit next to a disclaimer without turning into advice.
A prompt is not a plan
The release adds entry points called Featured: predefined prompts for specific jobs. The example is a wealth management plan based on the user’s assets and risk tolerance. Bybit says the reply comes back as structured analysis, with key data, clear insights, and suggested next steps. No sample is shown. Nothing in the announcement shows that those plans fit any particular person.
Hold the two claims together, because that is the product statement. The Featured prompt asks for a plan from assets and risk tolerance, and the reply is said to include next steps. In the same release, Bybit says outputs are for informational purposes only and do not constitute investment advice. The tension is on the page. It is also untested: no sample output and no outcome data have been published.
Bybit says instructions are executed in chats only after customer authorizations. The assistant can run against the main account, or the user can optionally deploy it in a dedicated Bybit AI sub-account. The release does not explain how that sub-account separates funds or risk. Until permissions and limits are defined, they are labels. A risk team cannot test them as a control model.
“Security-first architecture” is a design claim, with no controls, certifications, or technical notes attached. The same caution covers the broader positioning: deep global liquidity, robust security, transparent operations, and the assertion that Bybit is building the first intelligent platform connecting anyone, anywhere to the world’s finance. Those are the company’s words. So is the line that Bybit is trusted by more than 80 million users worldwide. That round number was already attached to the company in 2025 materials. Neither the release nor the HackerNoon piece offers a newer audited count.
On what the assistant can actually do, keep the two accounts apart. This release describes intended profile inputs and says execution waits on customer authorization. It does not list which actions a chat instruction can take. The earlier HackerNoon piece, which is sponsored, said the assistant could read plain-language intent and carry out actions, with examples that included checking a portfolio balance, placing a trade, and opening a support case. Neither account is independently verified.
For a risk committee, that unresolved gap is the story. Memory, as this release describes it, makes context sticky. Personalization makes the output feel bespoke. A disclaimer does not automatically stop a suggested next step from being read as a recommendation. Any firm putting an AI agent in front of a customer account is making the same trade: fewer menus, more context, more ways for a fluent answer to be mistaken for guidance.
Infrastructure, in a business that moves money, is the unglamorous list. What the system may see. How long memory lasts. Who can wipe it. Which actions need a fresh authorization. What “informational” means when the answer is built from open positions.
Questions worth asking before you call this an edge
What did the October release actually add?
Account-based personalization, long-term memory of preferences a user chooses to share, and Featured prompts.
Does Bybit AI give investment advice?
Bybit says no. Outputs are for informational purposes only and do not constitute investment advice. No sample plans or outcome data have been published to show how that line holds in a real reply.
Can it act on an account?
Bybit says instructions are executed in chats only after customer authorizations, on the main account or an optional Bybit AI sub-account. It does not say which actions qualify. Sponsored launch coverage gave trade placement and support-case creation as examples. Neither description has been independently checked.
Can users see, fix, or delete what it remembers?
The announcement does not say. It also does not name a retention period, a model provider, or how data is handled across borders.
Is “more than 80 million users” a new result?
Not on the evidence published. It is a company claim, and the same round number was already circulating in 2025 materials. Treat it as positioning, not as audited growth since then.
Why should a leader outside crypto care?
Because the pattern is an AI agent with customer state, memory, and a stated path to action, wrapped in a disclaimer. Bybit says further updates are coming as it builds toward AI-native financial services, with conversational intelligence as core infrastructure. The update that would matter is the operating detail this one skipped:
- whether users can view, correct, or delete memory, and how long it is kept
- a plain list of actions the assistant may take
- a permission model that states what limits actually block