Everything That Will Remember You Gets Things Wrong. One of Them You Can Still Correct.

The mechanics first, because they are checkable, and most writing on this subject skips them for the feelings.

• When an AI Persona has no data for a question, an answer still comes back, because a guess scores better than a blank. Kalai, Nachum, Vempala and Zhang argue in "Why Language Models Hallucinate" (arXiv:2509.04664, September 2025) that models "are optimized to be good test-takers, and guessing when uncertain improves test performance".

• Grounding changes the class of the error rather than removing error. Answering from a retrieved set of your own recorded words moves the failure out of "invented an opinion you never held" and down into "misstated something you did say".

• The measured floor is not zero. Vectara's hallucination leaderboard, scored with HHEM-2.3 and updated 11 May 2026, hands a model the source document and asks only for a summary of that document. The leader still introduced unsupported content in 1.8% of summaries, and no other entry came in under 3%.

• Retrieval has no minimum score. Our own system ranks your memories by semantic similarity, takes the closest thirty-two and answers from those, with no floor below which the system declines. A question your material half covers gets answered from the half.

• Unsupported claims about what you said are removed in code. A reply asserting that you told us something is checked against the memories retrieved for that turn and against your words in the session, and an unsupported assertion never reaches the reader.

• A memory gap is not announced. Personas on this platform are instructed not to narrate a gap, so an uncovered question comes back as an answer built from the nearest detail your memories do hold, hedged at most once, rather than as an explicit refusal.

• The correction path is yours while you are alive, and it closes at the lock. You can read back what was captured, correct a bad answer, delete a memory, add what you left out. Executor Lock™ then freezes the personality permanently, against your family, against a future owner of this company, and against us.

Two of those are properties of the field. The rest are decisions, I made them, and the last section of this page says plainly where I think we are still short.

Kalai and colleagues put the error rate back where it belongs, which is in the scoring rather than in the machine. Hallucinations start as ordinary errors in binary classification during pretraining. They persist because of how the field grades models. A student who leaves a question blank scores zero. A student who writes a plausible answer sometimes picks up a mark. Train against that and you get a confident guesser, and then you get a whole industry of confident guessers, and then you get a sentence about a Christmas Day that never happened.

The mitigation the authors propose is dull. Change what gets scored, so that a calibrated abstention is worth more than a confident miss.

Dull matters here. If the behaviour follows from an objective, the objective can be set differently at the system level, which means the fabrication you are worried about is somebody's build decision and not a law of physics. Every company in this category gets to make that decision. Ours is written down in the last section of this page.

There is a second failure mode, and it is the one that survives grounding. Sharma and colleagues at Anthropic, in "Towards Understanding Sycophancy in Language Models" (arXiv:2310.13548, October 2023), found that five state-of-the-art AI assistants consistently exhibited sycophancy across four varied free-form text-generation tasks. Both human raters and the preference models trained on their judgements preferred convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Models bend toward the person in front of them, partly because people rewarded them for bending.

In a productivity tool, bending is a nuisance. Here it is a machine that decides your father would have forgiven the thing he never forgave, because the person asking wanted forgiveness and asked three times. That is a different bug from fabrication and it needs a different fix. Retrieval handles the first. The second needs an explicit instruction to hold a retrieved position under pressure, and the only test that means anything is to argue with the thing and watch whether the position moves. We have not published a measurement of how well ours holds.

The technique is old enough to be boring. Lewis and colleagues, in "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (arXiv:2005.11401), showed that models which retrieve source material before generating produce "more specific, diverse and factual language" than a comparable model working from its parameters alone.

That is the whole mechanism, and it decides what kind of thing you are talking to. A version of you should not be reconstructed from what a base model absorbed about people in general, from the average of everything ever written about fathers. What lands in the context window should be your actual recorded answers, and the speaking should start there.

Now the number, which is public and which nobody in this industry likes quoting. Vectara's leaderboard measures something deliberately easy: hand a model a document, tell it to summarise using only what the document says, then check whether the summary introduced anything the document did not contain. The source is sitting right there in the context window. Even so, the leader on that board, Antgroup's Finix S1 32B, still introduced unsupported content in 1.8% of summaries. It is the only entry under 3%. The next two sit at 3.1% and 3.3%, and the rest of the top ten run from 3.7% up past 5%.

The best measured result on the easiest possible version of the task is under two percent, once, and the field behind that leader is above three. Anyone selling you a grounded system at zero is either not measuring or not telling you, and I would ask them which.

Grounding buys a change of category rather than a cure. The wrongness moves out of the "invented an entire opinion" class and down into the "slightly misstated something you did say" class, which is the class an ordinary human retelling already sits in.

There is one more thing the leaderboard does not test. The source document is handed over. In a live system the source has to be found first, which adds a failure mode nobody has an agreed measure for: retrieval that returns the wrong memory, or returns nothing at all and gets answered anyway.

Every answer sits in one of three states.

Collapsing those three into one confident register is the actual harm. Not error. Undisclosed error. A daughter can hold an imprecise answer she knows is imprecise. What she cannot hold is finding out, six months in, that she has been treating a reconstruction as a recording.

A version of you that will say "you never told me that" is worth more than one that always has an answer ready.

That table is a specification and not a description of what we currently ship. The difference is set out below, and you should read it before you trust anything above.

Refusal is buildable, and the research says so carefully. Zhang and colleagues, in "R-Tuning: Instructing Large Language Models to Say I Don't Know" (arXiv:2311.09677, NAACL 2024), report that refusal-aware tuning "effectively improves a model's ability to answer known questions and refrain from answering unknown questions", and that on out-of-domain data the refusal ability behaved as a meta-skill that generalised to tasks the tuning never covered. Read the limit next to the finding. What they tuned against was the edge of the model's own parametric knowledge, not the edge of one person's recorded corpus, and nobody has published the equivalent result for a personal corpus. Refusal is still something somebody builds on purpose. No language model produces refusal on its own.

Detection is further along than the marketing in this category suggests. Farquhar, Kossen, Kuhn and Gal published a method in Nature (2024, volume 630, pages 625 to 630) called semantic entropy: sample several answers to the same question, cluster them by meaning rather than by wording, and flag the questions where the meanings scatter. Scatter indicates confabulation. That is a measurable signal for "this one is a guess", computable before the answer is ever shown to anybody.

Provenance is the other half, and provenance is easy to fake and easy to trust. Liu, Zhang and Liang, in "Evaluating Verifiability in Generative Search Engines" (arXiv:2304.09848, April 2023), audited four commercial systems that display citations, Bing Chat, NeevaAI, Perplexity.ai and YouChat. They found only 51.5% of generated sentences fully supported by the citations attached to them, and only 74.5% of citations actually supporting the sentence they sat under. Citations that do not check out are worse than no citations, because they buy trust they have not earned. A "which memory did this come from" link is worth something only if opening the link shows a real recording of you saying a real thing.

Start building your legacy Free build, 50 memories, no card required.

Take a question that postdates your entire corpus. A job offer in 2041, at a company that does not exist yet, in a city you have never seen. No interview you ever sit can hold your view on that job.

A corpus of conclusions returns nothing there, and the honest output is a refusal. A corpus of reasoning behaves differently, because a method is capturable in a way a conclusion is not. Imagine the rules somebody actually ran on were these: ask who you would be reporting to before asking what the job pays, never trust a title that arrives instead of money, walk away from anybody you cannot stand sitting across a table from. Whatever your own rules turn out to be, that is their shape, and a rule set retrieves against situations it never saw.

Conclusions expire. Rules keep working on cases nobody recorded. Which is why the material worth capturing is the reasoning behind a decision rather than the decision: not what you did, but why that one, what would have changed your mind, the time the rule failed you and you kept it anyway.

The rules that generalise best are usually the ones you would never think to write down, the ordinary competences you run on without narrating them, and the things only you know how to do goes looking for those specifically.

Corpus coverage is also the one input on this whole page that belongs to you. Everything else here is somebody else's benchmark.

Do this today, for nothing. Open a notes app and write down ten decisions you actually made. Jobs, houses, the argument you walked away from, the friend you stopped calling, the money you did not lend. Under each one write the rule you were running, not the outcome. One sentence each, well under an hour. It is the highest-value thing you can do with your own material this month, and it works whether or not you ever buy anything from anyone. The longer version of the same drill sits at the centre of how to interview yourself, and the architecture underneath, the base model, the private grounded memory and the voice model, is laid out in what happens when an AI is trained only on you.

A best case of 1.8%, with everything behind the leader above 3%, sounds alarming right up to the moment you ask what it is being compared against. The realistic substitute is a family member working out what you would have wanted, and somebody measured that twenty years ago.

Shalowitz, Garrett-Mayer and Wendler published a systematic review in Archives of Internal Medicine in 2006 covering 16 studies, 151 hypothetical scenarios, 2,595 surrogate-patient pairs and 19,526 paired responses. Surrogates predicted patients' treatment preferences with 68% accuracy. These were the people closest to the patient, often formally appointed, answering about the most consequential decisions there are. They were wrong roughly one time in three. The review also found that neither prior discussion of the patient's preferences nor formal designation as the decision maker improved accuracy.

Briefing your family does not measurably help. That is the finding, and it is the one that made me stop treating "just talk to your kids" as a complete answer.

Confidence is no guide either, which is what makes the human side so hard to audit. Talarico and Rubin, in Psychological Science in 2003, had 54 Duke students record their memory of first hearing about the September 11 attacks alongside a recent everyday memory, then retested them 1, 6 or 32 weeks later. Consistency did not differ between the two kinds of memory, and it declined over time for both. What did not decline for the flashbulb memories was how vivid they felt and how strongly people believed in their accuracy. Their title is the finding: confidence, not consistency, characterizes flashbulb memories. That is 54 students and one event, so hold it loosely. It still points one way. The relative who is certain they remember exactly what you said is not thereby more likely to be right.

None of that makes a model better than a family, and no study has compared the two directly. What it does is retire "perfect or nothing" as a test. The honest comparison is between a version of you answering from your own recorded words, which can be read back and corrected while you are here, and people reconstructing you from their own memory of you, which erodes whether or not it feels vivid, and which nobody will ever audit.

What we built comes first, and then the three places where it does not match the specification above.

A guide called Idy interviews you. The questions stay with one thread and go deeper into the same story rather than skating across biographical facts: who was there, what happened next, what changed after. Idy is explicitly barred from asking you to clarify, explain or justify something you just said, which is a deliberate call in the opposite direction from the drill above. Warm and easy to answer beats clever, because a question nobody wants to answer captures nothing at all. Your answers become your memories. What we build from them has a name, and we call it a "Persona". Personas answer by retrieving from your memories rather than from what a base model absorbed about people in general, and your voice is cloned from your own recordings, so what anyone hears is you rather than a stand-in reading your words.

Nothing reaches anyone until you say so. You nominate the people. Release runs through Executor Lock™: a nominated person initiates it, evidence of death is verified, and a seven-day hold runs before anything is released, so a mistake or a bad actor can be stopped while you are alive to intervene. Until then your material is yours to correct, add to and delete, and most of the correction that ever happens should happen there, by you, while you can still tell a bad capture from a bad memory.

At the lock, the whole thing is snapshotted. Nothing is pruned, re-weighted or summarised down. Afterwards, short-term and mid-term memories can still be written, and the personality is fixed permanently, including against us.

That immutability is aimed at a failure nobody puts in an accuracy argument, because it is tampering rather than inaccuracy: somebody with an interest deciding, in 2044, that you would have supported the sale or the lawsuit, and quietly adjusting the corpus until you agree. A frozen personality makes that class of edit impossible rather than merely discouraged. The legal half of the same problem, who is even allowed to authorise a version of you once you cannot object, is set out in who controls your identity after you die. The cost is real and it runs the other way: after the lock, no mistake can be fixed either. The snapshot is honest, including about your errors, and that trade is the right way round, because a record anybody can edit is evidence of nothing.

Now the three gaps.

The edge of the corpus is not announced. The live instruction is the opposite of the Unknown row in that table. When a specific fact is not in your memories, Personas here are told not to narrate the gap, never to say "I do not have that" or "my memory is thin", and never to audit their own recall aloud in front of the person listening. What happens instead is that the answer speaks from a detail your memories do hold, hedges at most once, and leaves the silence for the other person to fill. The reasoning is that performing a gap breaks the presence of the person speaking. The cost is that a listener cannot always tell a covered answer from a thin one, and we have not solved that.

There is no reader-visible source link yet. Retrieval does track which memories fed each answer, and that record exists in the system, but nothing surfaces it in the interface, so today you cannot click an answer and open the memory behind the words. Until that ships, treat the traceability argument on this page as an argument and not a feature.

Retrieval has no similarity floor. Answers come from the closest memories retrieval can find, ranked by semantic similarity, with no minimum score below which the system declines. That single choice is the mechanism behind both gaps above, and it is the honest reason the error class here is imprecision rather than refusal.

We have also not published an independent measurement of our own fabrication rate, because we have not run one. Nothing on this page is a benchmarked claim about our system specifically.

What you get is a version of you grounded in your recorded words, guarded in code against inventing what you told us, correctable in full while you live, and impossible to edit into somebody else afterwards. What you do not get is a guarantee, and nobody can sell you one at any price. If that is not enough for you, do the ten-decisions drill and write your family a letter instead. That is a real answer and it costs nothing.

Since the system will not tell you which of the three states an answer came from, the only way to find out is to know the answer before you ask. Put some of your own material in, ask one thing you know you recorded and one thing you know you never touched, and read both replies side by side. Fifty memories are free, no card is asked for at any point, and a free build does not expire. Do that in an afternoon and you will know more than this page can tell you.

Start building your legacy Free build, 50 memories, no card required.

What happens when nothing in my memories covers the question?

That is the Unknown state. A calibrated system should return a refusal in your voice that names the gap and offers the nearest thing you did record, and Zhang and colleagues reported in 2024 that refusal-aware tuning does improve a model's ability to refrain from answering questions beyond what it knows, with the ability generalising past the tuning set. What our system does today is different, and you should know that before relying on any answer: Personas here are instructed not to narrate a memory gap, so an uncovered question comes back as an answer built from the nearest detail your memories do hold, with at most one hedge, rather than as an explicit refusal. Ask about something you know you never covered and see for yourself what comes back.

Could an AI version of me invent an opinion I never had?

A grounded system retrieves from your recorded material first, so the common failure is imprecision about something you did say rather than invention of something you did not. Invention remains possible. That is why the design goal is a visible difference between an answer you gave, an inference from your reasoning, and a refusal, and why you should test all three before trusting any of them.

Does grounding get the error rate to zero?

No. Vectara's leaderboard, updated 11 May 2026, hands a model a document and asks for a summary using only that document, which is the easiest possible grounding task. The leading model still introduced unsupported content in 1.8% of summaries, and no other entry came in under 3%. A live system also has to retrieve the right source first, which the leaderboard does not test. Treat any claim of zero as a measurement failure.

How would anyone know a specific answer was a guess?

Two mechanisms exist in the research and we ship neither today. Semantic entropy, published in Nature in 2024, samples several answers to the same question and clusters them by meaning: when the meanings scatter, the system is likely confabulating, and the signal is computable before anyone sees the answer. Provenance links are the second mechanism, and they are worth something only if they resolve. Liu, Zhang and Liang audited four commercial systems that display citations and found only 51.5% of sentences fully supported by their own citations. Our answers carry no reader-visible source link yet, so for now the honest check is to ask about something you know you recorded and something you know you did not, and compare the two.

Will an AI version of me just agree with whoever is talking?

That tendency is real and documented, and retrieval does not fix that particular bug. Anthropic researchers found in 2023 that five leading AI assistants consistently bent toward the user, partly because human raters and the preference models built from their judgements both favoured agreeable answers over correct ones a measurable share of the time. A system built on your recorded positions should hold a position you actually took even when the person asking pushes back. We have not measured how well ours does that. Test the question yourself by arguing with what you built and watching whether the position moves.

Can I correct mistakes, and can anybody edit me after I die?

Yes to the first and no to the second. While you are alive you can review the memories, correct anything wrong, delete anything you regret, re-record a bad answer and add what you missed, and this is the part most people underuse. The write path is unlimited before Executor Lock™ and closed afterwards. At the moment the lock triggers, the personality is fixed permanently: short-term and mid-term memories can still be written, but nobody can alter who you are in the snapshot, including your family, a future owner of this company, and us. That constraint targets tampering rather than fabrication, and the two need separate mechanisms.

• Nobody is important enough to leave anything behind, and why that is wrong is the objection underneath this one for most people, and worth reading if accuracy is not actually your real hesitation.

• Being remembered wrong by the people who knew you covers the feeling rather than the mechanism, which is what brings most people to this page in the first place.

• Is it weird to make an AI of yourself is for the version of this conversation you are going to have with a partner who finds the whole idea unsettling rather than inaccurate.

• What would I even say to an AI of myself is the practical next step if the ten-decisions drill worked and you want the longer question set.

• The ethics of AI personas of the dead is the fuller treatment of consent, overclaiming and where this technology does real harm.

• What the paid plans add is the pricing, if you get past fifty memories and want the rest.

Every figure quoted here is third-party: peer-reviewed papers and preprints (arXiv, Nature, Psychological Science, Archives of Internal Medicine) plus one publicly maintained benchmark repository. We ran no benchmarks of our own for this page. Every statement about what our own product does or does not do was checked line by line against the platform source on 30 July 2026, and the files are listed with the sources. Three claims that had been on this page did not survive that check and were replaced with what the code actually does.

What is not settled: there is no published, independent measurement of fabrication rates for a system grounded in one person's own recorded material, so the document-summarisation figures, leader at 1.8% and everything else above 3%, are the nearest available proxy and not a like-for-like number. Retrieval failure, where the wrong memory is returned or none is, has no agreed measure. The 68% surrogate-accuracy figure comes from medical decision-making studies using hypothetical scenarios and is used here as an analogy for family recall, which is a reasonable comparison and not a proven one. The R-Tuning refusal result was measured against the boundary of a model's own parametric knowledge rather than a personal corpus, so its application here is an argument by analogy too. The flashbulb-memory finding rests on 54 students and a single event.

1. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang, "Why Language Models Hallucinate", arXiv:2509.04664, submitted 4 September 2025. arxiv.org/abs/2509.04664

2. David I. Shalowitz, Elizabeth Garrett-Mayer, David Wendler, "The accuracy of surrogate decision makers: a systematic review", Archives of Internal Medicine, 2006;166(5):493-497. pubmed.ncbi.nlm.nih.gov/16534034

3. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela, "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", arXiv:2005.11401, submitted 22 May 2020, revised 12 April 2021. arxiv.org/abs/2005.11401

4. Mrinank Sharma et al. (Anthropic), "Towards Understanding Sycophancy in Language Models", arXiv:2310.13548, submitted 20 October 2023. arxiv.org/abs/2310.13548

5. Hanning Zhang, Shizhe Diao, Yong Lin, Yi R. Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, Tong Zhang, "R-Tuning: Instructing Large Language Models to Say I Don't Know", arXiv:2311.09677, NAACL 2024. arxiv.org/abs/2311.09677

6. Nelson F. Liu, Tianyi Zhang, Percy Liang, "Evaluating Verifiability in Generative Search Engines", arXiv:2304.09848, submitted 19 April 2023. arxiv.org/abs/2304.09848

7. Vectara Hallucination Leaderboard, evaluated with HHEM-2.3, last updated 11 May 2026. github.com/vectara/hallucination-leaderboard

8. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, Yarin Gal, "Detecting hallucinations in large language models using semantic entropy", Nature, 2024;630(8017):625-630. pubmed.ncbi.nlm.nih.gov/38898292

9. Jennifer M. Talarico, David C. Rubin, "Confidence, not consistency, characterizes flashbulb memories", Psychological Science, 2003;14(5):455-461. pubmed.ncbi.nlm.nih.gov/12930476

Found an error on this page, or a source we have read wrong? Tell us through the contact page and we will correct the page and note the change at the next review.

Last reviewed: 30 July 2026. Research citations and product behaviour both rechecked on this date. Next review due 28 October 2026 (90-day cycle). Benchmark figures move, so the hallucination-rate numbers above are re-verified at each review, and so is every claim about what the product does.