Fictional records · every answer cites its source line

The report changed. The answer shouldn't be a guess.

Talk to your records. Voice Memory checks corrections, conflicting values and missing approvals, then shows exactly what the evidence supports.

Ask the records

Illustrative synthetic records only. Not a diagnostic or clinically validated product.

ENWhat's my ferritin?

ListeningReading 2 reports for SYN-0001Amendment found · 15 OctUsing the amended value

Preliminary · 13 Oct 5 ng/mL
Amended · 15 Oct 50 ng/mL
a-2026-10-amended.pdf · p.1

Amended 15 October 2026: ferritin result corrected from 5 to 50 (transcription error).

Corrected value “Your ferritin is 50 ng/mL. The 12 October result was corrected from 5 to 50, and I'm using the corrected value.”

Real engine output on fictional records, in the language each question was asked. Try it below.

The 80-second film

People read their records before anyone explains them.

68.6% of US adults with online records opened their test results before hearing from a clinician. JAMA Network Open, Aug 2026. HINTS 2024, n = 4,982 1 in 3 of them said they understood those results only fairly well or poorly. Same study: 26.8% fairly well, 6.6% poorly 40M+ people a day ask ChatGPT about health. ECRI ranks misuse of AI chatbots the #1 health-technology hazard of 2026. ECRI Top 10 Health Technology Hazards, Jan 2026

Records change: a preliminary result gets amended, a lab switches units, a proposal gets revised. A chatbot that reads the newest line, or the wrong patient's, sounds confident either way.

Ask the records. Out loud or typed.

Pick Lab reports, Northstar deal or Your documents, then ask. The source line opens beside each answer.

The record decides factual answers. Voice Memory may clarify what you mean, but it only states what the documents support.

Voice calling is temporarily off; typed asking is fully live.

Live Bolna web call · uses your microphone

Pick a question, type or dictate your own.

Answers come from the documents only, checked against the exact line they cite.

    Source
    The original page opens here, with the cited line marked.

      Bolna hears it. The record decides it.

      Speech in, a checked sentence out, and the source on screen while you listen. Voice calling is paused for now; typed asking above is fully live.

      1. 01

        You ask, in your words

        Speak or type in English, Hindi or Hinglish. Fillers and “no, sorry” corrections are cleaned, and test names and IDs are spelled from your case. Phone calling is paused; dictation and typing are live.

      2. 02

        It works out what you mean

        “My” becomes your case, “bayalis hazaar” becomes $42,000, “it” becomes the thing you just asked about. Advice it can't give, it says so.

      3. 03

        The record decides

        The original line, then its status: which patient, which unit, amended or preliminary, proposed or approved.

      4. 04

        A friend answers, with the line

        One warm sentence in your language, every word lit as it is spoken, and the cited line open on screen.

      What it catches

      Six things a fluent assistant gets wrong. Click one to ask it live.

      Memory graph

      Every value is tied to its source sentence, and to what replaced it.

      Reports, values and the exact sentences behind them, joined by corrects, supersedes and not comparable edges - rebuilt from the current documents on every view, the same evidence the engine answers from.

      Open the memory graph

      Put a verdict behind your agent's answers

      Your voice or chat agent already talks. Voice Memory decides which version of a fact is in force, refuses when the record can't say, and returns the line it rests on, in English, Hindi or Hinglish.

      Support lines on changing terms

      Rate cards, policy wordings and fee schedules are revised. The caller gets the current figure, never the draft, with the clause it comes from.

      Report callbacks

      "Did my value change?" across amended reports, refusing when units differ. Health use waits for clinician review.

      Deal desks and agencies

      "What did we last propose, and was it approved?" across proposals, minutes and email, with the open items listed.

      // remote: add as a custom connector, nothing to install
https://voice-memory-beige.vercel.app/api/public/mcp

// or local, over stdio
{ "mcpServers": { "voice-memory": {
    "command": "node", "args": ["voice-memory/mcp/server.mjs"] } } }
// remote tools: ask_records, list_cases, create_workspace, workspace_status

      Pilots: 30 days, up to 5 of your documents, your own agent. Priced per verified answer; abstentions are free.

      Request a pilot Why this, and the numbers

      Where it stands

      Built for the Bolna Symphony Buildathon, running on fictional records.

      Works now

      • Typed questions on this page, straight from the evidence (voice calling is paused)
      • Amended, preliminary, proposed, superseded and still-open values, tracked per value
      • Lab results compared across reports, refusing when units or patients differ
      • Short spoken answers built from the record, with the cited line on the original page
      • PDF, Word, scanned image and text ingestion with page-anchored citations
      • Lab explanations in English, Hindi or Hinglish, in the language you ask
      • A memory graph of documents, values and the sentences behind them
      • Your own documents on this page, in a private workspace deleted after 30 idle minutes
      • A key-free API for scripts and agents (llms.txt, OpenAPI)
      • Benchmarked on outside-authored records: real meeting transcripts (QMSum) and SEC contracts (CUAD)

      Next

      • Reworded questions: meaning-based search lifts the right line to first place on outside contracts (44% → 50%), but a "closest line" pointer stays off until it is right 9 times in 10 (measured)
      • Clinician review of lab explanations before any patient pilot
      • Phone-number calls and WhatsApp voice notes through Bolna

      Ask out loud. Check the line.

      Ask the records See the memory graph