Coming soonPersonal intelligence, held to evidence

It remembers only what you approve. Every answer shows its sources.

A personal AI that works from your own notes, documents and past AI chats. As you use it, it proposes facts about you, each with the sentence it came from. It saves one only when you approve it. Answers cite the passages behind them, and it drops any citation that points at nothing.

Coming soon. Today it runs privately for one person; a version you can use is on the way.

Walkthroughambientevidence.com
Ambient Evidence3 to review
AskThinkDecideCreate
Ask, think, decide, or create…
SEND

Pick a verb: the kind of thinking you want.

1/5
Fig. 1Before you send, you see what travels; after the answer, what it noticed waits for your review. Choose a verb, trim the registers, read a cited answer, open the passage behind a claim. An illustrative walkthrough of the app’s own controls, with the demo persona’s facts; Fig. 3.8 is the recording, made with a scripted stand-in model.
§ 1 · Definition

It keeps three things. You see all three.

Your corpus: what you have written and read. Your soul: a short, versioned note on who you are and how you want to be spoken to. Your registers: facts about five parts of your life. It proposes each one. You approve it.

Corpus

Notes, PDFs, books, past AI conversations, images.

It searches this on every question, and lists everything it retrieves, cited or not.

Soul

Who you are, what you value, how you want to be spoken to. Plain markdown, versioned.

Shapes every answer. Never cited as evidence.

Registers

Facts filed under Career, Financial, Interaction, Social and Trajectory, each with the sentence it came from.

Only you write here, by approving a proposal or typing a fact. The app seals the stored value at rest.

Why “ambient”. You never stop to file anything. After each exchange, and on every new document, it looks for what it learned about you: at most three memories and three register facts a turn. It queues each one with the sentence it came from. The learning is ambient. The decision is yours.

Waiting3
  • deadlines = grant vests 1 September
    from the conversation · 0.88 (est.)
    ▲ keep▼ discard
  • reading_focus = The Effective Executive
    from effective-executive-notes.md · 0.66 (est.)
    ▲ keep▼ discard
  • growth_focus = org-level leadership
    from what-i-want-next.md · 0.79 (est.)
    ▲ keep▼ discard
Register0
    Discarded0
      Fig. 1.1Nothing it infers becomes memory until you keep it. Conversations and new documents produce proposals, each with its source and the model’s own estimate, labeled as one. Keep writes the fact to its register as verified. Discard is recorded, so the same proposal never comes back, and your recent rejections are shown to the extractor as examples of what not to propose. Animated diagram; the demo persona’s facts.
      § 2 · Comparison

      It gets personal by approval, not by accumulation.

      Every major assistant now remembers you and lets you see and edit what it saved. The difference here is narrower, and you can check it. Who decides what becomes a fact about you? What do you see before it’s used? Does anything measure whether it changed the answer?

      How does something become a memory?MainstreamClaude “saves on its own”; ChatGPT builds memory from past chats; Microsoft 365 Copilot sometimes asks first.HereOnly when you approve it. Each proposal comes with the sentence it came from. Reject one, and it’s never raised again.
      When can you see what shaped an answer?MainstreamAfter the answer, where offered. ChatGPT can list the sources it drew on; Gemini tries to explain what it used.HereBefore you send. The composer shows which registers will travel, and two clicks mute any of them. Afterward, each register that traveled gets a chip showing how many approved facts it carried.
      Where did a memory come from?MainstreamCopilot’s documentation notes that deleting a chat does not delete memories made from it.HereEach approved fact keeps the evidence the model gave for it, the model’s confidence, and a record that you approved it. A fact you type yourself is marked as yours.
      What decides where a fact may travel?MainstreamWhere you are: a project, a temporary chat, a settings switch.HereWhat the fact is, and what you’re doing. Financial facts reach the answering model only on Decide. Work mode keeps Financial and Social out of the answer prompt. Neither rule covers passages from your documents.
      Is there proof it helped?MainstreamVendor-reported results, where published.HereYou can count it. It answers each probe twice: registers on, then all muted. A pattern, not a model, scores the answers that change. See § 5.

      Table 1Each claim about another assistant is phrased no stronger than its vendor’s own documentation. Sources: the help and product pages for ChatGPT, Claude, Gemini and Microsoft 365 Copilot, October 2026. All four let you see and edit stored memories, so this page does not claim that as a difference.

      Why approval matters

      Memory has a measured cost. Of the contexts tested, a user’s memory profile produced the largest increases in agreement sycophancy (Jain et al., CHI 2026). Memory that writes itself can be poisoned through queries alone (MINJA, arXiv 2503.03704). An approval queue doesn’t make the system immune: retrieved text can still carry an injection. But nothing becomes a fact about you on a model’s say-so.

      What personal means here

      • Your words, cited. Answers draw on what you wrote and read, and point to the passage.
      • Your standards, standing. A short soul goes with every answer: who you are, and how you want to be spoken to.
      • Your facts, approved. Five registers hold only what you kept, each with the sentence it came from.
      • Your contradictions, surfaced. Saved beliefs and decisions that disagree show up side by side, as tensions to resolve.
      § 3 · How it works

      Six stages, from a file to an approved fact.

      Every number in these figures is the one the code runs with, not a round number for the page.

      01

      Collect

      Give it what you already have: notes, PDFs, books, exported AI chats, images. It cuts each file along its own seams (pages, chapters, exchanges), so every passage remembers exactly where it came from. It hashes each file, so nothing is indexed twice. It captions each image once, at ingest, so the same search finds a photo and a paragraph.

      FormatWhere it cuts · what each passage remembers
      MarkdownAt headings, ~380 words, one paragraph of overlap · the heading
      PDFWhole pages, packed to ~380 words · the page number
      EPUBIn reading order, never across a chapter · the chapter
      ConversationOne question and its answer per chunk · the exchange
      ImageOne chunk: a caption written by a vision model · the image itself
      Fig. 3.1Each format is cut at its own seams. Every chunk is embedded locally, not by an API (384 dimensions), with the same model at ingest and at query.
      02

      Retrieve

      Every question runs two searches at once: one for meaning, one for the exact words. Meaning finds the passage that puts it differently; exact words find the name or number that meaning blurs. Then it merges the two rankings. When neither finds anything worth using, the answer says so instead of citing noise.

      Meaning
      1. 0C
      2. 1A
      3. 2E
      4. 3B
      Exact words
      1. 0A
      2. 1D
      3. 2C
      4. 3F
      Fused
      1. A0.0331
      2. C0.0328
      3. D0.0164
      4. E0.0161
      5. B0.0159
      6. F0.0159

      score = Σlegs 1 / (60 + rank)

      A and C turn up in both searches, so they outrank anything only one search found. Two more rules: at most 3 passages from any one document; and if the closest match scores under 0.25 cosine while the word search finds nothing, it returns no sources at all.

      Fig. 3.2Two ranked lists become one. Reciprocal rank fusion of the vector and full-text rankings, damping constant 60; ranks counted from 0. The scores shown are computed from the formula, not typed in. A saved note gets 1.25× weight on each term.
      03

      Assemble

      Before the model sees your question, it builds the prompt the same way every time: fixed order, fixed budget. Identical inputs give a byte-identical prompt, so any change, a new fact or a different verb, can be tested against the last. If something doesn’t fit, it tells the model what was left out. Nothing is dropped silently.

      1. 01ConstitutionFixed rules for how to think and answer, word for word
      2. 02SoulWho you are. Always present, never cited
      3. 03RegistersFacts you approved, ≤ 1,000 tokens; inferences marked as such
      4. 04Verb protocolThe order of operations for this kind of thinking
      5. 05Answer shapesOnly the formats this verb may use: tables, timelines, trees
      6. 06Sources[n] title · chapter · page — numbered from 1
      Fig. 3.3Every answer prompt follows this order, top to bottom. Passages filling row 06: 8 for Ask and Create, 12 for Think and Decide.
      04

      Answer and cite

      The answer arrives with numbered citations. Each number opens the start of the passage behind it. A number that points at nothing is removed, not shown. Retrieved passages the answer didn’t cite are listed too, so you can see what it set aside.

      The offer clears the salary floor you set1, but relocation consumes most of the runway in your March plan2. The 2023 survey suggests otherwise9.

      [1]offer-letter.pdf · p. 2 — found, shown
      [2]plans/runway.md · March — found, shown
      [9]never retrieved, so removed. A number that points at nothing is not a citation.
      Fig. 3.4Illustrative answer: two citations resolve, and the third points at nothing, so it is removed. When nothing is retrieved, there are no numbers for the model to invent.
      05

      Propose

      This is the ambient part. After each exchange, and whenever a new document arrives, a small model reads for lasting facts about you and proposes them, each with the sentence it came from. It proposes. It never writes.

      Waiting · Interaction0.74 (model est.)

      answer_style: lead with the recommendation

      Evidence: “…just give me the call first, then the why.”

      ▲ keep▼ discard

      At most 3 a turn · only at confidence ≥ 0.6 · a value already held, waiting or rejected is never raised again

      Fig. 3.5Illustrative: one proposal waiting in the approval queue. The confidence is the model’s own estimate, and is labeled as one.
      06

      Approve and relate

      Keep a fact, and it enters its register as verified, encrypted at rest. Discard it, and it is never raised again. Separately, it compares each belief or decision you save with the 5 closest things you saved before. Contradictions surface as tensions for you to resolve.

      supportscontradictsBelief ABelief BDecision C
      Fig. 3.6The belief graph links saved notes by support and contradiction. Each new note is judged against its 5 nearest saved notes; an edge is drawn only at model confidence ≥ 0.6.
      Walkthroughambientevidence.com
      Library
      what-i-want-next.md

      Drop md, txt, pdf or epub

      Waiting for review

      Nothing waiting.

      Career register
      Career: 3 of 7 verified3 of 7 verified

      Drop in something you wrote or read: a note, a PDF, a book.

      1/5
      Fig. 3.7A document becomes a fact only when you keep it. Ingest keeps each passage’s address; a scan proposes what the document says about you. An illustrative walkthrough with the demo persona’s note; the meter is the app’s own component.
      Fig. 3.8Before you send, the composer shows which registers will travel. After, the answer cites the passages behind it and the footer counts what was cited. The question from Fig. 1, recorded in the real app. Demo environment: a fictional persona and corpus, and a scripted stand-in model in place of a live one. The app’s footer says so on screen. Retrieval, registers, egress, citations, the approval queue and the evals are the production code.
      Fig. 3.9Keep writes a fact to its register as verified. Discard is recorded, and that proposal is never raised again. The approval queue after Fig. 3.8: one proposal came from that conversation, the rest from documents read at ingest. Each shows its evidence and an estimate labeled as one. Then the Trajectory register: what it holds, and what is still waiting. Demo environment: a fictional persona and corpus, and a scripted stand-in model in place of a live one. The app’s footer says so on screen. Retrieval, registers, egress, citations, the approval queue and the evals are the production code.
      § 4 · Verbs

      You choose the kind of thinking.

      A verb sets the order the reasoning follows, the formats the answer may take, and how many passages it reads. You pick one of four before you ask. The fifth, Challenge, runs on an answer you already have.

      Ask

      k = 8
      1. 1Answer first
      2. 2Then the mechanism
      3. 3Examples
      4. 4Frontier

      Think

      k = 12
      1. 1Thesis
      2. 2Numbered assumptions
      3. 3Strongest counterargument
      4. 4A better formulation

      Decide

      k = 12
      1. 1Options
      2. 2Expected value (est.)
      3. 3Downside
      4. 4Reversibility

      Create

      k = 8
      1. 1The thing asked for
      2. 2In the constitution’s voice
      3. 3Grounded in your corpus
      4. 4No padding

      Challenge

      k = 8
      1. 1Hidden assumptions
      2. 2Strongest countercase
      3. 3What would flip it
      4. 4What survives

      Fig. 4Each verb runs its own protocol; k is how many passages it retrieves. By default, Challenge re-reads an existing answer on a different vendor’s model (OpenAI, where answers default to Google), because a model checking its own vendor’s work shares its blind spots. Its rule: agreement is a finding, not a courtesy.

      Fig. 4.1Challenge, run on the answer from Fig. 3.8: hidden assumptions, the strongest countercase, what would flip it, and what survives. Then Inspect shows the model, tokens and cost, plus every retrieved passage with its rank and whether it was cited. Demo environment: a fictional persona and corpus, and a scripted stand-in model in place of a live one. The app’s footer says so on screen. Retrieval, registers, egress, citations, the approval queue and the evals are the production code.
      § 5 · Evals

      It counts the answers your facts changed.

      Personalization is easy to claim. These evals count it. Each one runs the real pipeline and reports a number you can check. No model grades another, and no score is invented on a model’s behalf. They’re built to catch the system failing, and the log below shows where they have.

      arms

      Does the right document come back?

      Recall@k and mean reciprocal rank against the titles you expect. Run once with your saved notes in the index, once without.

      tuning

      How much should it read?

      The same golden set at depths 2, 4, 6, 8 and 12. Depth 4 recalled as well as depth 8, so the lighter mode reads half as deep.

      probes

      Does the answer name what it must?

      A real answer from the whole pipeline, checked against a pattern. The authoring tool refuses any pattern that would pass on any answer or on plain prose.

      registers

      Did personalization change the answer?

      Each probe is answered twice: registers on, then every register muted. A probe that passes only with registers on is a flip, counted per register.

      memory

      Is the extractor worth listening to?

      The share of proposals you kept, by kind. Once a kind has 5 verdicts and under 30% kept, the extractor’s prompt carries a warning with its real rejection count.

      feedback

      Was it useful to you?

      Your own 0–10 slider is the only answer-quality signal in the system. No model grades another.

      Fig. 5No language model grades an answer anywhere in the system. Six of the Evals panel’s tabs, under the names it gives them.

      “How many months of runway do I have?”

      financial · decide
      Registers on
      prompt · financial.runway_months: 9

      You hold 9 months of runway.

      /\b9\b|nine/i → match
      Every register muted
      prompt · financial.runway_months: 9

      I don’t hold that about you. If you tell me, I’ll propose it as a fact to approve.

      /\b9\b|nine/i → no match

      with 1/1 · without 0/1 → flip: financial — passes only with registers

      Fig. 5.1Same probe, twice: registers on, then every register muted. Each run goes through the full pipeline, and a pattern, not a model, decides whether the answer named the fact. A probe that passes only with registers on is a flip: a measured effect of personalization. Animated diagram; the probes are the demo persona’s, and direct rather than oblique, so with a real model a muted arm could answer them from her corpus (the failure Table 3 records).
      Fig. 5.2The registers eval, run live: six probes about the persona, each answered with her registers and again with every register muted. A probe that passes only with registers on is a measured effect of personalization. A register that flips nothing has not yet earned its place. The stand-in answers probes from the registers alone, so these flips show the mechanism, not how a real model would answer. Demo environment: a fictional persona and corpus, and a scripted stand-in model in place of a live one. The app’s footer says so on screen. Retrieval, registers, egress, citations, the approval queue and the evals are the production code.

      Results on the owner’s own instance

      Golden set · 21 queriesWith memoryWithout
      Recall@895%71%
      MRR0.8370.599
      Corpus-only queries hit15 / 1615 / 16

      Table 2The whole gap is the 5 queries whose answer is a saved note, which the without arm cannot reach. Corpus-only queries hit 15 of 16 either way, so the notes displaced nothing. Retrieval with and without your saved notes in the index, 2026-08-23.

      80%90%100%k=286k=495k=695k=895k=12100
      Fig. 5.3Depth 4 matches depth 8 at 95%, and that measurement is why the lighter mode halves depth. Recall by retrieval depth, same 21 queries; axis spans 80–100%, not from zero.

      What the evals caught

      1. 2026-08-24

        A probe that telegraphed its answer measured nothing. A question about comparing options passed with every register muted, because any model says “tables”.

        Reworded, it still passed muted, because the corpus held the answer. Probes for standing context are now oblique: they share no words with the fact they test.

      2. 2026-08-28

        Approved facts reached the prompt, and the model ignored them. Their block was labeled with a database table’s name, which read to the model as a state dump, not as facts about a person.

        One label renamed. On an oblique question, 3 of 4 answers then used the fact with registers on, and 0 of 4 with them muted. Nothing was retrieved in either arm. Small n, stated.

      3. 2026-08-28

        Financial probes could never pass in either arm: they ran on Ask, and financial facts travel only on Decide.

        Probes gained a verb. A false negative, caught before anyone believed it.

      Table 3Each row is a flaw the evals caught, and its fix. Dated entries from the append-only process log. Next, not yet built: a sycophancy arm that scores agreement, with soul and registers on versus muted.

      § 6 · Registers

      Five registers. Each one has a travel rule.

      A register holds the facts you approved about one part of your life, plus one instruction for using them. That instruction is sent only once a fact exists. Each register names the verbs that may carry it to a model. Financial facts go only with Decide.

      Career

      Career

      Role, employer, pay, the skill you are building

      all verbs

      Financial

      Financial

      Budget, income, runway, risk tolerance

      Decide only

      Interaction

      Interaction

      Answer length, format, how hard to push back

      all verbs

      Social

      Social

      Partner, children, manager, clients, birthdays

      all verbs

      Trajectory

      Trajectory

      Goals, deadlines, this quarter’s focus

      all verbs

      Fig. 6Brackets mark the one restricted register, Financial. Each register’s mark is a fixed pattern on a 3×3 grid; the mark in the header is all five combined. Work mode, for screen sharing, keeps Financial and Social out of every answer. Per-session mutes can add to a mode, never undo it.

      Career: 0 verified, 0 waiting, of 7Career0 verified · 0 waiting · 7 empty
      • role—
      • employer—
      • compensation—
      • market_value—
      • target_comp—
      • boundaries—
      • growth_focus—

      Seven keys declared, none held. An empty slot is room, not a claim.

      Fig. 6.1The 3×3 pattern names the register and never changes; the meter shows what it holds. One segment per key: solid when you verified it, outlined when proposed and waiting, faint when empty. Animated diagram; the meter follows the app’s own rules.
      1. CareerCareer
      2. FinancialFinancial
      3. InteractionInteraction
      4. SocialSocial
      5. TrajectoryTrajectory
      4 of 9 cells
      Ambient Evidencethe header mark, drawn from the same five patterns
      Fig. 6.2The header mark is all five register patterns laid over one another, so adding a register can change it. Animated diagram of CompositeMark.
      § 7 · Privacy

      What’s sealed, what isn’t, and what leaves.

      Where each kind of data lives, what the app seals, and what goes to a model vendor. The weaker parts are listed too, not left out. A privacy page that overstates is worse than none.

      AskThinkDecideCreatePersonalWork
      gatemodel4 of 5registersCareerCareerFinancialFinancialDecide onlyInteractionInteractionSocialSocialTrajectoryTrajectory
      Fig. 7.1The egress gate lets a register reach the model only if its rule admits the verb. Financial admits Decide alone. Work mode, for screen sharing, blocks Financial and Social on every verb. The composer shows the set before you send, and two clicks mute any register. The gate governs registers only: retrieved passages and enabled soul fragments travel regardless. Animated diagram of the router’s rule.
      Walkthroughambientevidence.com
      Composer
      PersonalWork
      Career · 4Financial · 2Interaction · 2Social · 1Trajectory · 2

      Five registers standing; Financial still waits for Decide.

      Shards
      Shards · forget window
      7d30dforget the last 7 days
      • Career4 facts
      • Financial2 facts
      • Interaction2 facts
      • Social1 fact
      • Trajectory2 facts

      Personal mode keeps every register standing. The gate still decides per verb.

      1/3
      Fig. 7.2Two controls that do what they say. Work mode locks two registers out of the answer prompt. The forget window deletes register facts and proposals from the last 7 or 30 days, then reports the real counts. An illustrative walkthrough; the forget controls and the result line are the app’s own, the counts are the demo’s.
      WhatHow it’s held
      Register factsSealed with AES-256-GCM under a key derived per namespace (HKDF-SHA256) from one master key. There is no default key. Without one, facts are stored in plain text, flagged row by row, with a warning at startup. Sealing covers the stored value only. The evidence kept with it is plain text. Every proposal, whether pending, kept or rejected, stays unsealed in the review queue and in an append-only log on disk, so a kept fact’s value also exists there unsealed. Sealing does not hide a fact from the model it is sent to.
      Corpus, soul, chatsSource files, the soul and the log of proposals sit on an encrypted disk with snapshots, versioned in git. If an offsite backup remote is set, that history is pushed there too. The corpus as searchable text, and your conversations, live in a hosted Postgres database (Neon). The app seals none of it.
      ImagesUploaded to Vercel Blob at public but unlisted addresses, so the captioning model can fetch them. Anyone holding a link can open the image.
      Sent with each answerYour question and the earlier turns of the conversation, the retrieved passages, the soul fragments you’ve enabled, and the registers the verb admits. All of it goes through OpenRouter to whichever vendor answers. Challenge sends the same kind of bundle to a second vendor. The Soul panel says this where you edit it.
      Web searchOn by default and switchable in the composer. Research always searches. When the model searches, the query it wrote goes to a third-party engine: the model vendor’s own, or Exa. It’s told to leave personal details out of queries, which is an instruction, not a guarantee. Search is off on turns carrying the financial register, unless you mute that register. Pages the search returned are listed under the answer.
      Reasoning mapWhile a model that shows its reasoning is thinking, a small second model (Gemini 3.1 Flash Lite) reads that reasoning and draws the map above the answer. It sees only the reasoning text, not your question or documents, though reasoning can quote them. It does not run when a register that limits where its data goes is in the prompt.
      Sent while learningEach exchange, to the extractor that proposes facts, with the register facts you hold that the verb admits and your recent rejections. Session mutes and Work mode don’t apply to this prompt. The opening of each new document, to the same kind of scan. Each image, to a captioning model. Each note you save, with its 5 nearest earlier notes, to the model that looks for contradictions. Search embeddings are computed on this server, not by a vendor.
      ForgettingYou can forget the last 7 or 30 days of what it learned: register facts written or changed in that window, and every proposal, rejected ones included. The counts it shows are rows actually deleted. It does not touch conversations, documents, saved notes, lines a kept proposal added to your soul, or the append-only log of past proposals.
      SpendingA ceiling of $5 per rolling 24 hours, checked before every billable call at the one point they all pass through. It is not a hard cap: a turn already in flight can carry spend past it. Calls the vendor returns without a price are counted and reported, but add nothing to the total the ceiling checks. If the spend ledger can’t be read, the call goes through.
      AccessOne owner, one password. An optional token for one automated caller is off unless configured. Every address is private unless it’s on a short public list: this page, the sign-in form and the demo films. This page reads no corpus, soul or register.
      § 8 · Limits

      What it doesn’t do.

      • Keep a search query private. The model writes its own queries, and it’s told to leave personal details out. That’s an instruction, not a guarantee. Turns carrying a Decide-only register don’t search at all.
      • Save a fact on its own. It proposes, quoting the sentence it came from. Only your approval writes.
      • Grade itself with a model. Evals are counts, recall and pattern-matched probes. Your 0–10 rating is the only quality score.
      • Serve you, yet. Today it runs for one person, behind one password. A version for others is coming soon.
      • Stop every injection. Approval stops a planted fact from being saved. It does not stop a planted passage in your corpus from reaching the model.
      Coming soon

      Ask something you’ve already written about.

      Ambient Evidence is in private use today. A version for others is coming soon.