What generic AI doesn't know is you

How It Works.

From one record to insight

vivome structures records into 12 blocks and runs them through 30+ analysis engines for cross-domain insight.

Fact computation is separated from text generation

Most AI chatbots hand the whole question to a language model and let it generate an answer. vivome adds a step before that. Factual queries, “when did I last visit a clinic,” “how many headaches this month,” are computed by a deterministic lookup that scans and aggregates the event records on your device, not by the language model. Dates, frequencies, and statistics always come from your actual records; the model only adds the phrasing on top. Because fact computation and text generation are separated, there is no path for the model to invent a date.

How generic AI works, and its limits

Generic AI is a language model trained on public text. It generates the statistically most likely next token given the input, so without access to your personal data it answers questions about a specific individual with general averages or estimates. ChatGPT's memory stores summarized preferences and facts, such as “likes Sapiens.” It is not a structured time-series with timestamps and intensity, so it cannot reconstruct an aggregate like “12 headaches in the last three months,” and items the model infers from chats are not visible to the user as a full list. Without grounding data, questions about dates or frequencies can produce values that don't match reality (hallucination). Fine-tuning on human feedback (RLHF) can make a model agree with the user (sycophancy). In April 2025, OpenAI rolled back a GPT-4o update for this. vivome AI works differently. A lookup such as “when did I last visit a clinic” is computed by aggregating on-device records directly, not by the language model, and generated answers are grounded only in the structured records you've entered.

Only as much as the question, from a structured record

Records in vivome are normalized into events with a code, a timestamp, and an intensity, not free text. When a generated answer is needed, the system assembles a data inventory from only the domains relevant to the question, activity and daily logs for an exercise question, and leaves unrelated domains out of the prompt. Even then, only display labels are passed, not internal codes. If there is no record for the period, it states “no records,” so the model can't fill the gap with a guess.

Three gates before anything reaches an external model

Before any external language-model call, a request passes three gates. (1) A content gate: if a sensitive topic, a breakup, trauma, self-harm, is detected, it is handled only on-device and never sent to the cloud. (2) A consent gate: external AI is off by default (fail-closed), and no call happens before explicit consent. (3) Masking: if a request is sent, sensitive references, people, memories, records, are replaced with placeholders first. The API key lives only in a server-side proxy, so the model can't reach your account or database, only what the app placed into that one request.

Output passes deterministic post-processing

Before a response reaches the screen, it goes through deterministic post-processing. Internal codes become labels, expressions that flatter or simply agree with the user are removed (empathy like “that must have been hard” is kept), and sentences that assert causation are rewritten with lower certainty. Diagnoses, scores, and commands like “you should…” are suppressed both in the system prompt and in this post-processing. The result stops at “what is observed in the records,” not “what you should do.”