Baxnet Ideas · Founder note
The Models Will Change. The Person They Need to Understand Won’t.
TL;DR: AI models are becoming easier to replace. The difficult thing to recreate is an accurate history of the person they are meant to represent.
Ask an AI to plan a family trip and the conversation often begins with a small autobiography.
There are two children. One room will not work. A late flight is fine on the way out but miserable on the way home. One person needs gluten-free food. Another would rather spend more on a direct flight and less on the hotel. The last holiday looked good on paper and failed because the transfer was too long.
None of this is difficult to explain once. The odd part is having to explain it again to the next assistant, the next travel service and whichever model replaces both of them a year later.
AI is becoming extraordinarily capable. It still begins many useful tasks knowing remarkably little about the person it is meant to help.
That is the product gap I think Mimoto can fill.
General intelligence still needs personal context
A model can know a great deal about hotels, flights, nutrition and family travel without knowing which trade-offs matter to one particular family. It can produce a good answer to the general question and still miss the best answer for the person asking it.
The missing material already exists, but in fragments. It sits in messages, calendars, purchases, notes, financial records, previous applications and the reasons people gave when they changed their minds. Different companies build narrow profiles from those fragments for their own purposes. The individual rarely has an equivalent intelligence layer working on their side.
Personal Intelligence is my name for that layer: a persistent understanding of a person built from their history, preferences, relationships, goals, constraints, decisions and changing circumstances.
The useful object is not the largest possible archive. It is a compact, current and inspectable model that can help answer questions such as: What has this person tried before? Why did they reject the cheaper option? Which constraint is temporary? What happened after the last decision?
Research into search personalisation found that long-term, across-session history could be more useful than short-term session behaviour, although the value varied by user and query. More recent personalised-memory benchmarks reach a similarly qualified conclusion: linking events over time can improve preference understanding, but even advanced systems still struggle when preferences change or old information should be forgotten. The history helps only when the system understands which parts remain relevant. (Management Science, PERMA)
Conversations are the beginning, not the boundary
Mimoto starts with private conversations because they contain an unusually dense record of a life.
People explain decisions in messages. They discuss purchases before they appear in a transaction history. They reveal who matters, which plan keeps being postponed, why a previous option failed and how a relationship changed. Much of that context never reaches a formal profile.
Today, Mimoto turns supported iMessage history on Mac and user-selected WhatsApp exports on iPhone into local reports, patterns, timelines and structured exports. That is the current product. It helps a person inspect conversation history that is otherwise trapped inside search and scroll.
The longer-term direction is broader. Mimoto can move from understanding conversations, to understanding the person behind them, to connecting decisions with their reasons and outcomes. Eventually, with explicit permission and a published integration path, a user-chosen AI could ask for the small piece of context required for a task.
That final step is a direction, not a feature we claim to offer today.
A decision is more useful than a preference label
“Ben likes direct flights” is useful. It is also thin.
A better record might say that Ben chose a direct flight over a cheaper connection for a short family trip because arrival time mattered more than price, and that the decision worked well enough to change what he is willing to pay next time.
That structure preserves more than a preference. It connects the situation, options, constraint, decision, reason and outcome. It can also record that the preference later changed.
This matters because people can state their current preferences tomorrow. They may not be able to reconstruct why they made a decision five years ago, what alternatives they considered or what happened afterwards. Some historical context becomes impossible to recreate once the messages, notes and memories behind it disappear.
More history does not automatically mean more value. Old details become irrelevant. Some beliefs become wrong. A system that remembers everything without understanding change may become confidently personal and consistently outdated. The job is selective preservation, correction and forgetting.
Models should be replaceable
Someone may use one assistant today, another next year and a model that does not yet exist after that. Rebuilding the person inside every provider creates repeated effort and a new form of lock-in.
The better architecture separates the model from the personal context.
Personal Intelligence should be portable enough to survive a change of model, inspectable enough to correct, and governed well enough that access can be revoked. Important claims should carry timestamps, confidence and provenance. An external agent should receive the minimum useful context for its task, not a permanent copy of the conversations from which that context was learned.
That idea is becoming less hypothetical at the infrastructure level. The W3C’s Verifiable Credentials work provides standard building blocks for machine-verifiable claims and selective disclosure. The EU Data Act gives users new rights to access and share data from connected products. Neither creates Personal Intelligence by itself, but both point towards a world in which useful context can move under clearer user authority. (W3C, European Commission)
The value is in what the intelligence can do
I do not think the strongest economic case for Personal Intelligence is that somebody will pay a large sum to buy a person’s raw data.
The more credible value appears when better context improves a decision: avoiding an unsuitable purchase, finding a better service, noticing a recurring cost, reducing search time, remembering an important constraint or helping an agent negotiate from the buyer’s side.
This distinction becomes more important as AI moves from answering to acting. The Agentic Commerce Protocol already gives compatible agents and merchants a way to construct and complete a checkout, with payment credentials limited to a particular amount and seller. Transaction infrastructure is appearing. The open question is whose interests the agent will represent when it uses the context around that transaction. (OpenAI, Stripe)
A seller can use personalisation to increase conversion or charge more effectively. A buyer-aligned agent could use Personal Intelligence to compare suppliers, preserve the person’s constraints and refuse an attractive offer that is wrong for them. The same information can create very different outcomes depending on who controls the system and what it is paid to optimise.
That is why the individual has to remain the customer.
The bet still has to earn its way into the present
There is a serious counterargument to all of this.
Models may learn enough about a person from a short conversation. Large platforms may control the agents people actually use. Consumers may prefer convenience to portability. Rich history may add less value than expected, while creating a much larger security problem if it is wrong or compromised.
Mimoto cannot rely on a future infrastructure thesis to excuse a weak product today. Conversation reports, recall, structured exports and relationship context need to be useful now. Portability and agent access should extend that value later rather than justify collecting information with no present purpose.
Still, the timing matters. A better model can be adopted when it arrives. A forgotten reason, vanished conversation or changing relationship cannot always be recovered on demand.
The models will change. The agents will change. The person they need to understand remains the same person, even as that person changes too.
Mimoto’s ambition is to preserve the part that matters, keep it under the individual’s control and make whichever AI they choose better at acting in their interests.