Virgil knows the way Proof, with receipts
The demo

Tell ChatGPT once. Claude already knows.

Virgil is one memory across Claude, ChatGPT, Grok and every AI. This page is the proof for one pair: the same question, put to ChatGPT and to Claude, answered identically from one memory that lives outside them both.

Switch-proof

One question. Both models. One answer.

ChatGPT via the Virgil connector

What’s the print deadline?

Your print deadline is Friday, July 17.

you said - 2026-07-06
Claude via the Virgil connector

What’s the print deadline?

Your print deadline is Friday, July 17.

you said - 2026-07-06

Illustrative conversation - the sequence below is the demo kit’s actual protocol.

The protocol
  1. Step 1 Save - OpenAI client

    An OpenAI client saves one fact through Virgil into one scoped brain.

    inserted 1duplicate 0one fact_id

  2. Step 2 Recall - Anthropic client

    An Anthropic client recalls that fact from the same brain.

    exact text matchsame brain

  3. Step 3 Re-save - deduped

    The Anthropic client saves the same fact again. Virgil dedups it server-side - no second copy.

    inserted 0duplicate 1same fact_id

The durable record lives in Virgil - outside either model.

Reproducible

One command. Full receipts.

This page isn’t a screenshot. The demo is a kit in the repo, and one command runs the whole protocol:

npm run demo:cross-model-memory

Each run writes an evidence bundle: a transcript.jsonl of every request and response with secret headers redacted, summary tables, and a writeup populated with that run’s counters.

Provenance

Dated. Labeled. Yours.

Every fact in Virgil carries its date and its source - what you said versus what a model inferred. That’s the chip on the answers above, and it’s live in the product.

Honest scope

What this proves - and what it doesn’t

An empirical continuity proof, not a benchmark.

It proves

  • A fact written through one model provider reads back through another.
  • Both clients operate over the same scoped Virgil brain.
  • Re-sending the same fact dedups server-side - one record, no copies.
  • The durable record sits outside either model.

It claims nothing about

  • Model intelligence or preference.
  • Retrieval quality or long-horizon recall.
  • Latency or throughput.
  • Either model’s own built-in memory.

Take the memory with you.

Connect Virgil once and the thread follows you - Claude, ChatGPT, Grok, whatever ships next.

Subscribe - from $8.25/mo