I want personal AI to learn the shape of my life

A personal AI should eventually know more about the situation than I can fit into a prompt.

My work, commitments and constraints don’t reset with a new conversation. I want a continuing system that carries that context forward and anticipates things I haven’t thought to ask.

Human–AI symbiosis means the person and system could learn together. Remembering my favourite response format is only a beginning.

A record of my life is not yet a model of it

A specific expectation gives the system something to test. “This work block may slip” is more useful than an untestable label like “you always overcommit.”

Make the expectation testable

1. 01

   Write it before the event.

   A forecast needs a specific expectation that can turn out to be wrong.

   Illustrative work block · not a personal record

   Expected

   Work block

   Observed

   Not known yet

   Record the expectation first

   “Given the commitments already in the calendar, this block is likely to slip.”

2. 02

   Put the outcome beside it.

   Collecting records becomes useful when they can test what the system expected.

   Illustrative work block · not a personal record

   Expected

   Work block

   Observed

   Started later

   Compare against what happened

   A mismatch can be inspected. It does not become a personality label.

3. 03

   Correct the model, not the person.

   Inspect the mismatch. A changed goal or a deliberate choice is not necessarily a failure to follow the plan.

   Illustrative work block · not a personal record

   Expected

   Work block

   Observed

   Started later

   Revise the assumption

   Was the estimate wrong, did a dependency change, or did the person choose a different goal?

Illustrative planner, not personal records or measured accuracy. A useful forecast must beat a simple baseline and improve a real decision.

OwnTracks could supply location context; calendars record intentions, banking records describe cash-flow timing, and nutrition logs could add another perspective.

Those are possible roles, not a claim that every feed is connected. Collecting more data can feel like progress without answering a useful question.

The test is whether a forecast recorded beforehand can beat a simple routine-based guess—and whether acting on it helps enough to justify the attention.

The world model has to include the person who can change it

By a personal world model, I mean a working account of goals, facts, constraints, dependencies and expectations. I don’t mean a complete simulation of a human being.

The personChoose or change the goal

Correct a fact. Do something unusual. Decide that the old objective no longer matters.

1. 01ObserveRelevant, current context
2. 02ForecastBefore the outcome
3. 03ChooseKeep the person’s agency
4. 04CompareOutcome against expectation

↶ Corrections revise the next expectation

A proposed learning loop. Forecasting components exist; the complete path to measured personal benefit has not been demonstrated.

I can change a goal, choose something unusual or correct an event. A model that predicts yesterday’s habits while dragging me back towards them would be a poor partner for becoming different.

The work block started on time. That observation fits either of these stories:

The warning helped

The person moved the preceding commitment after seeing the forecast. The original collision may have been avoided.

The warning was unnecessary

Nothing changed, and the preceding commitment finished early. The forecast may have overstated the risk.

One on-time start cannot tell us which explanation is right. Keep the action and its timing alongside the result; repeated comparisons are needed to assess whether warnings help.

An invented example. Keep the original forecast, the person's action and its timing. Forecast accuracy and the benefit of acting require separate evidence.

Where DNA fits, and where it doesn’t

I am interested in biological information, including my genomic data, where there is a justified use. That is an ambition—not evidence that my DNA is ingested or a genetic model is working.

Location

Where a routine or transition happened

Calendar

What was intended, not proof it happened

Money

When commitments and cash flows occur

Nutrition

A possible input; not a verified integration here

Genomic data

Only for a justified, appropriately interpreted question; not a definition of a person

Source roles are design intentions, not a claim that these feeds are all connected. Genomic ingestion is an ambition, not a completed integration.

The National Human Genome Research Institute explains that a polygenic risk score is only one factor in predicting disease. A measurement relevant to one question is not a discovery of who someone “really is.”

The system should explain why information belongs in a decision. Private records can stay private while we discuss the design.

What exists today

My repository contains location-transition and cash-flow forecasting components, a private forecasting library and an encrypted ledger. They are pieces of engineering, not evidence that the complete partnership works.

The broader prediction model is still recorded as disabled pending a verified source adapter. MyFitnessPal is a possible input, not a verified integration. The complete observation-to-decision-to-benefit loop has not been demonstrated.

JustSwipe is the interface idea for decisions emerging from this work. OpenClaw is the intended continuing operator. The names do not establish completion.

Symbiosis should expand what a person can do

In “Man-Computer Symbiosis”, J. C. R. Licklider described close cooperation, including help with formulating problems.

What could one person attempt with a system that retains context, searches possibilities and learns from outcomes alongside them?

That is the future I want to test. Knowing more about me should increase my ability to act, rather than make the advice harder to question.

Sources and further discussion

• Wu and colleagues: LongMemEval (ICLR 2025). This benchmark tests memory across conversations, including changes over time, updated information and knowing when to abstain. Those are useful requirements for a personal system. Passing a memory benchmark would still not prove that its forecasts are sound or that its advice improves someone’s life.