Proactive personal AI · Research prototype
An assistant that predicts before it acts and learns from the gap.
MeApp aims to spot problems before you ask, prepare a way to solve them, anticipate needs you haven't thought of yet, and remember context so you never have to explain the same thing twice.
The idea
Expect first, then measure the difference.
MeApp is built on one simple idea: before the AI starts a task, it independently predicts what will happen. Afterwards, it compares the real outcome with that expectation. Every measurement meets in a single shared numeric metric, and that metric drives what the assistant pays attention to, what it remembers, which method it trusts, what it learns, and how it changes its behaviour.
Notice before being asked
Spot problems before the user raises them, and prepare how to solve them.
Anticipate unspoken needs
Foresee what the user will need next, even before it occurs to them.
Keep continuity
Remember the past so the user never has to explain the same thing twice.
How it works · a worked example
One experience, several behaviours improved.
- Predict. Before writing a report, MeApp predicts it can produce a complete result from the information at hand.
- Seal the expectation. The prediction is produced separately from the process doing the work, and recorded before any result is seen.
- Act. During the work, it turns out a critical piece of data is missing.
- Compare. The gap between expectation and outcome is reflected in the shared metric.
What that one signal changes
The same measurement feeds several parts of the system at once:
The core research problem
Making the measurement trustworthy.
The real research question is making this measurement reliable. If the measurement is wrong, what the system learns and how it adapts can be steered in the wrong direction.
Judging how surprising an outcome is requires reading the context and cross-checking with more than one calculation. Different domains bring their own appropriate checks, and all of them meet in one common measure.
The approach is inspired by the prediction-and-adaptation idea of the Free Energy Principle (FEP), treated here as a software simulation rather than a biological claim.
Layered checks, one score
- Code
- Deterministic checks enforce explicit limits and hard facts.
- Model
- A separate model evaluates the parts that need interpretation.
- Domain
- Each domain contributes checks suited to it.
- Output
- All of them combine into a single surprise score.
The whole loop at a glance
From one task to a better next task.
The predictor and the worker never share context, so the expectation can't be bent to fit the result. Everything downstream (attention, memory, trust, behaviour and Dream) reads the same single score.
Architecture
Three layers: physiology, psychology, soul.
The design is organised in three layers. Each principle in the top layer is tied to behaviour through code, and its usefulness is measured. Where interpretation is needed a model evaluates; where limits are explicit, code enforces them.
Physiology
Measurement and feedback: recording expectations, collecting outcomes, and computing the shared surprise metric.
Psychology
Skills and behaviour: the methods the assistant uses, how they are chosen, and how they change with experience.
Soul
Priorities. Nine principles decide what matters before, during, and after a task.
Sets the expectation a task begins with.
Sets what knowledge the task begins with.
Asks questions under uncertainty.
Chooses the fitting step and output true to fact.
Changes a method that isn't working.
Tries new paths within limits.
Protects the user's time and attention.
Looks ahead.
Turns experience into skill.
Learning over time
Dream: connecting past experiences.
Dream is the part of the system that links experiences together. It goes back over individual episodes and looks at similar failures and better-than-expected results side by side.
Its proposals are tested first, and permanent changes are made under human control. Over time, this accumulated experience should help the assistant recognise the early signs of an approaching problem and prepare a solution that has worked before.
Dream, skill learning, and day-to-day behaviour are all tied together by the same shared metric.
Questions Dream asks
- Which assumption was wrong?
- Which method worked unexpectedly well?
- What should we change next time?
Built on Claude
Focus on the core, not the plumbing.
The first prototype runs directly on the Anthropic API with the Claude Agent SDK, while MeApp's prediction and measurement core is developed separately. Ready-made agent infrastructure means isolated contexts, tool handling, and permission management don't have to be rebuilt.
Expectation in, outcome out
Record the expectation before a tool runs and collect the result after it.
Separate prediction and evaluation
Prediction and evaluation run in separate contexts, so the predictor never sees the result.
Reuse what has been proven
Methods that pass testing are stored and reused when needed.
Human approval where it matters
Rules decide which steps proceed automatically and which need the user's approval.
Status & plan
Where MeApp stands today.
- Stage
- Research and architecture design; first prototype in progress
- Company
- Onflowly · bootstrapped, not yet incorporated
- Founder
- Muzaffer Cihad Öner · independent developer, Turkey
- Stack
- Anthropic API · Claude Agent SDK · Python
- Availability
- Not publicly available yet
Next step: test the key assumption
The first step is to test this approach with a small, observable prototype. That means generating predictions across many different scenarios, evaluating outcomes, and running the loop again and again.
Above all, many comparative experiments are needed to see whether the shared metric actually improves later decisions.