Eli Life-Capture Lab
I pointed a local AI system at my entire life.
Continuous audio. Every useful sensor I can access. As much context as I can responsibly capture from the device I already carry.
I am using myself as the test subject to answer a question that is both technically interesting and deeply uncomfortable:
What becomes possible when an intelligence system has a dense, continuous record of a real person’s life?
The Eli Life-Capture Lab is testing which products, workflows, and human outcomes emerge from radical personal capture, and which ideas collapse once the data stops feeling magical.
Everything in the experiment remains local. My life is not being sent to cloud models so someone else can learn from it.
More data does not automatically create more intelligence.
The basic assumption behind personal AI is that more context will produce better answers.
Give the system your conversations, schedule, location, movement, relationships, work, health signals, and decision history. Once it knows enough, the thinking goes, it will finally understand your life.
That may be true.
It may also produce an enormous archive of repeated routines, unfinished sentences, accidental correlations, background noise, and moments that matter only because a sensor happened to capture them.
The technical ability to record a life does not prove that the recording contains a useful product.
The Lab begins where the collection ends.
The research problem is compression.
Capturing information is becoming easier.
Microphones can listen. Cameras can observe. Phones already detect movement, location, orientation, activity, and patterns of use. A system can collect far more about a day than a person could ever review manually.
The harder problem is deciding what any of it means.
A useful system would need to distinguish between a passing comment and a commitment, between a repeated habit and a meaningful change, and between two events that merely happened close together and two events that are actually connected.
It would need to compress hours of ordinary life into the small amount of context that could improve a future decision.
That is the real experiment.
Not whether we can record everything, but whether anything valuable survives after the noise is removed.
The setup is intentionally extreme.
This is not a proposal that everyone should begin recording their entire life.
I am pushing the experiment as far as I reasonably can because mild capture will not reveal the full shape of the problem.
If a system only receives the thoughts I deliberately type into it, then it sees the version of my life I already decided was important. That can be useful, but it cannot tell us whether continuous context reveals something I would not have known to preserve.
The Lab therefore tests a more radical setup:
- Capture as much potentially useful context as the device can provide.
- Keep the processing and memory local.
- Use one willing test subject: me.
The discomfort is part of the research. Any product built from this work should have to earn the right to exist under the strongest version of the question.
A life is not a transcript.
Speech is only one layer of context.
A transcript may preserve the words from a conversation while losing who was present, what happened beforehand, whether I was exhausted, what decision followed, and whether the conversation changed anything afterward.
A useful life-capture system cannot simply create a searchable wall of text.
It needs to understand events.
An event might include what was said, where it happened, which people or projects were involved, what changed, and whether the moment should influence anything in the future.
That requires bringing together multiple weak signals without pretending that any one of them provides the full truth.
The research is about whether those signals can become a meaningful representation of a life rather than a larger surveillance log.
The system has to know what deserves to survive.
Most moments should probably be forgotten.
The sound of a door closing, another walk through the kitchen, or half a sentence overheard from a television does not become valuable simply because it entered a database.
The system needs a reason to retain something.
That reason may be that the moment contains a commitment, changes an earlier belief, reveals a recurring problem, affects a relationship, introduces new evidence, or connects to an unresolved decision.
Even then, the system may not need to preserve the original recording forever.
It may be enough to retain a concise memory with a clear connection to the event that created it.
The Lab is testing how much can be discarded while preserving the context that could still matter later.
The hypothesis
The central hypothesis is that continuous local capture can produce forms of persistent intelligence that deliberate note-taking alone cannot provide.
That does not mean the raw data is the product.
The product would come from a pipeline that turns experience into usable context.
01
Capture
02
Events
03
Meaning
04
Memory
05
Support
Capture the experience
Gather the available signals without assuming that every signal is meaningful.
Identify events
Separate a continuous stream into moments that may have their own purpose, participants, subject, or outcome.
Interpret the meaning
Determine what changed, what was promised, what remains unresolved, and why the moment may matter.
Build selective memory
Preserve the useful context while discarding information that adds volume without improving understanding.
Bring it forward later
Reconnect the memory to a future decision, conversation, relationship, or pattern when it becomes relevant.
The experiment is whether that final step becomes meaningfully better because the earlier capture was dense.
What could become useful?
The Lab is not beginning with a fixed product specification.
It is looking for outcomes that repeatedly emerge from the data and prove useful in real life.
One possibility is better decision memory. Eli may preserve why I made a choice, recognize when the assumptions behind it have changed, and bring the earlier reasoning forward when I need to revisit it.
Another is commitment tracking. A promise made casually in conversation may be easy to forget even though it matters to another person.
A third is pattern recognition. The system may notice that a recurring frustration, idea, or concern appears across contexts that I would never have manually connected.
Those possibilities sound compelling in theory.
The Lab exists to determine whether they remain compelling after the system has to find them inside an actual life.
A correlation is not an insight.
Dense personal data makes it easy to discover patterns.
It also makes it easy to invent them.
A system might notice that two things often happen on the same day and produce a convincing explanation for why they are connected. With enough signals, it will always be possible to find something that appears meaningful.
That is dangerous.
A personal intelligence system should not quietly turn weak correlations into beliefs about the person it serves.
The Lab needs ways to distinguish between:
- A repeated observation.
- A plausible interpretation.
- A conclusion supported by enough evidence to influence a decision.
Uncertainty must remain visible.
The system should be able to say that it noticed something without pretending it understands why it happened.
The person must be able to correct the record.
A captured event is not the same as the truth.
Audio can be misunderstood. Context can be missing. A joke may be treated as a commitment. A difficult moment may be given more weight than years of ordinary behavior.
The system will make mistakes.
I need to be able to inspect what Eli inferred, correct the meaning, remove the memory, or tell the system that an event should not influence anything later.
That correction should become part of the intelligence.
A system that records a person without allowing that person to challenge the interpretation is not persistent intelligence.
It is an automated witness that believes itself.
Local processing changes what can be explored.
The information in this experiment may include private conversations, unfinished thoughts, sensitive work, personal relationships, and moments that were never intended to become part of any company’s dataset.
Sending that context to cloud models would make the experiment easier.
It would also undermine the reason for doing it.
The Lab uses local models and local storage because the intelligence developed from a person’s life should remain with that person.
That constraint will limit what the system can do. Local models may be slower, less capable, and more difficult to run within the battery, memory, and performance limits of an iPhone.
Those limitations are part of the research.
The question is whether useful intelligence can be created without making privacy someone else’s promise.
Privacy is necessary, but it does not make the experiment harmless.
Keeping the data local reduces an important class of risk.
It does not remove every ethical problem.
Other people may appear in my conversations and environment even though they are not the subject of the experiment. Continuous capture can affect the behavior of the people around it. A system may preserve information that another person reasonably expected would disappear.
The Lab has to treat consent, visibility, retention, and deletion as part of the architecture.
“Local” cannot become an excuse to stop asking whether something should be captured at all.
This should not become surveillance sold as self-knowledge.
There is a version of this idea where every interaction is scored, every behavior is interpreted, and every deviation becomes a notification.
That product may be technically impressive.
I do not believe it would make a person more free.
A useful system should help the person understand what matters without turning ordinary life into a performance for an invisible observer. It should not optimize the person toward goals they never chose or make them anxious about every inconsistency.
The Lab is looking for intelligence that supports agency.
If the system makes the person feel managed by their own data, the product has failed.
What would prove the hypothesis wrong?
The experiment may reveal that radical capture is unnecessary.
Deliberate notes, selected recordings, and direct conversations with Eli may provide nearly all the useful context with far less noise, technical cost, and ethical risk.
Continuous capture may consume too much battery, create too many false memories, or require so much correction that the person becomes an unpaid data curator.
The system may find interesting patterns that do not improve a single real decision.
The hypothesis should weaken if:
- Dense capture does not outperform intentional capture in useful ways.
- The cost of filtering the noise exceeds the value of the signal.
- The privacy and consent risks cannot be reduced enough to justify the product.
A result that says “do not build this” would be valuable.
Some ideas should fail in the Lab before they are placed around another person’s life.
The benchmark is not how much Eli remembers.
A system that stores more data will always appear more capable in a memory demonstration.
Ask it what restaurant I visited four months ago, and a complete archive may produce the answer.
That is not enough.
The research should be judged by whether the captured context helps with outcomes that matter.
Did it recover a commitment I would otherwise have broken?
Did it bring forward evidence that changed a decision?
Did it reveal a meaningful pattern that survived inspection?
Did it help me understand something about my work, relationships, or behavior that I could act on responsibly?
The size of the memory is not the measure of intelligence.
The value of what it helps me do next is.
What the Lab will publish
The work should make both the usefulness and the discomfort visible.
That includes the capture methods, local architecture, event-detection experiments, compression approaches, false correlations, missed moments, and examples where a proposed feature turned out to be surveillance theater.
The Lab should also compare continuous capture with less invasive alternatives.
If a five-minute reflection at the end of the day produces better memory than twenty-four hours of sensor data, that is the result worth publishing.
The goal is not to prove that maximal capture wins.
It is to find the minimum amount of capture required to create something genuinely useful.
The connection to Eli
Eli is fully local persistent intelligence for iPhone.
The life-capture Lab explores one extreme version of the context Eli could use. It tests whether continuous experience can be compressed into memory that helps a person make better decisions without surrendering their life to the cloud.
Not everything explored in the Lab should become part of Eli.
Some forms of capture may be too invasive, too noisy, or too expensive to justify. Other experiments may reveal smaller, more deliberate workflows that provide most of the value.
The Lab has permission to push the idea too far.
Eli has to become a product a person can trust.
FAQ
Frequently asked questions
Why use continuous capture instead of journaling?
That is one of the questions the Lab is testing.
Continuous capture may reveal important context that a person would not know to record. It may also prove less useful than deliberate reflection.
Is this the same as Eli?
No.
Eli is the product. The life-capture Lab is an experiment that tests one of the most extreme possible sources of context for persistent intelligence.
Will life capture become an Eli feature?
I do not have a verified answer yet.
The Lab exists to determine which parts are useful, safe, and worthy of becoming products.
Is any of the data sent to cloud AI models?
No. The defining constraint of the experiment is that the processing and memory remain local.
The ability to capture a life does not prove that we should.
The valuable question is what the system can help a person understand, remember, or decide that they could not have done as well without it.
I am using my own life to test that question under a strict local-only constraint.
Some of the data may become useful intelligence.
Most of it may deserve to disappear.
The Eli Life-Capture Lab is where I intend to find the difference.
Follow the research.
Get updates when experiments ship, fail, or change what we believe about the architecture.