Your AI has Amnesia.
MnemosyneC gives any AI you already use (ChatGPT, Claude, Gemini, or any model) a permanent, private memory that lives on your own computer, with a source you can check for every answer.
Dr. MnemosyneC has the Cure: Substrate. Because DORA (Derive Once, Remember Always) has the Map.
Together they give your AI a persistent context with no ceiling we’ve found: 5.32 billion tokens of stored records, each reachable by address, and a 50-token pointer to the stored record, with sources you can check at will.
Three proofs. Run them yourself.
A question answered once, by anyone, never has to be paid for again. One of Us = ALL of US.
- In your browser, right now: mnemosynec.org/prove/. 300 answers recalled from the record, each one hash checked live while you watch. Zero model calls.
- In your browser: mnemosynec.org/asked/, "Somebody Else Asked." Open a question somebody else already asked and published, and the answer comes back from the record instead of being paid for again, with your browser checking the card against its own address. Zero model calls. Publish your own at mnemosynec.org/asked/publish/.
- In your browser: mnemosynec.org/hunt/, the Scavenger Hunt, the Haystack Needle. Give an address, get back the exact card: your tab loads the 999 published cards, checks every hash, and hunts for needles by address, one of them made up so it has to come back NOT FOUND. Zero model calls. The Founder's own machine receipt: 67 needles asked for by address in a haystack of 660,682 files, 67 found, exact to the byte, about 50 tokens each, in 0.048 milliseconds each.
Join the cooperative, $5 a year
The free Conduit · about 39 MB · no install. Unzip it, double-click Start Conduit, and when Windows warns choose More info, then Run anyway. It runs on your machine only and shows you real answers with real sources, at no cost per answer.
The same 300-card proof, run inside your own browser tab · nothing installed, nothing sent to a model · your browser computes the hash and shows it next to the one we published.
Want the whole app instead of just the proof? Download the signed Windows installer (about 92 MB, Windows only for now).
See the receipts →, the four measurement receipts behind every number on this page, as plain pages a person or an AI assistant can read straight through.
Corpus size.
5.32 BILLION TOKENS OF STORED RECORDS, EACH REACHABLE BY ADDRESS. So far.
About 5,000 times what the best AI can hold at once. The price stayed the same from a thousand files to 660,000. Here’s the receipt.
What a replay costs versus asking again.
A 50-TOKEN POINTER TO THE STORED RECORD. Every time.
A 50-token pointer opens the exact record (CHEAP). Asked for by address: 67 needles out of 67, found in a haystack of 660,682 files, exact to the byte (GOOD), in 0.048 milliseconds each (FAST).
Cheap. Good. Fast. Free. Pick four.
Download it. Prove it. Post it. The first 500 checked posts become Early Adopters.
Exact replay of stored cards (zero model calls after the mint).
ANSWERS ONE MACHINE PAID FOR, REPLAYED ON ANOTHER THAT NEVER ASKED. 300 for 300.
Free. Fast. Sourced. And this is just the Beginning...
After ANYONE on the Substrate learns an answer, EVERYONE has it: traced to its record every time, over 99% faster, at zero additional AI cost. Not an answer key: a taught method, run as code, that gives the same answer every time.
Sharing is off until you turn it on. Your card store stays on your computer; a prompt you send to an online AI is processed by that provider.
Read more: What the 300 of 300 is, exactly
What the 300 of 300 is, exactly. Those Hollerith Cards were built for those exact 300 real MMLU-Pro questions, and 16 of the 300 were repaired with the answer key. It is exact recall (the same stored method replayed by code, on a second machine that never saw the key) not blind solving of new questions. Read the receipt →
Somebody already paid for that answer. Now it’s yours to keep. The more you use it, the better it gets.
The bigger and slower your AI, the more a remembered answer saves.
Model accuracy, replayed and blind.
| Prepared-card replay, exact recall | 300 of 300 (16 repaired with the answer key) |
| 100 unseen questions, first pick, blind | 83 of 100 |
| Same 100, after the follow-up loop, not blind | 90 of 100 |
Next: 700 more MMLU-Pro questions. Our prediction of the result is already sealed and hash-stamped, so you can check that we called it before we ran it.
Disclosed in the sealed envelope: 80 of the 700 questions had already been minted when it was sealed.
Verification details
Sealed prediction sha256 4551cb9a377385b15c867f419cb20a0a88691da301cb580a7476458321694c1d, hashed 2026-09-16T20:39:38Z, public at mnemosynec.org/sealed-prediction-2026-09-16/, OpenTimestamps-stamped and archived at web.archive.org the same night.
PROVE IT for yourself in your first session. Free. No ads. No account. Map your part of the world’s data. Join the Future: Today. One of Us = ALL of US.
I have a video that I made for the Original Plan to make a Decentralized Factory for Home Based Manufacturing, starting with the HexIsle Game that uses my Tereno System, planned for Christmas 2025. And then I changed it right after making that, when all the federal programs started to be cut, and created the 16 Initiatives. It shows what I was trying to do for the Decentralized Factory, and it gives BruckLyn' a lot of credit, showing his permission as the song plays ... that is what I pivoted away from to make the 16 Initiatives, and thus developed MnemosyneC.
Music: Moonshot by BRUCK'LYN, used with permission
What your first session looks like
- Tell it something. “I’m working on a report about water treatment costs in rural areas.”
- Come back three days later. Start a fresh session. Ask about it. It remembers.
No copy-paste. No re-explaining. Pick up exactly where you left off, every time.
Free. Or pair MnemosyneC with a paid AI like Claude or GPT for faster, bigger answers. Either way, the substrate carries the memory, so your paid AI bill stays smaller too.
How it works
1. Ask.
Use the AI you choose. MnemosyneC keeps the answer with its sources, so you, and anyone on the Substrate, never have to ask that question again. That is DORA: Derive Once, Remember Always.
2. Trace.
Click the source on any answer in the Substrate Record and follow it all the way down the chain to the original. No source, no trust.
3. Keep.
The next time anyone asks, it replays from the record for about 50 tokens. Think a source is wrong? Add a better one. The record keeps both: the current answer and its history. Read the full mechanics →
Why a cooperative?
90% of all 28 filed provisional patents, with over 2,200 innovations and 2,473 claims behind them, and more to come, are pledged to the cooperative under its charter: yours to use and build on as a member, never sold for extraction. See the Pledge →
Help Each Other Help Ourselves. And we’ll go first.
Continued on the Cooperative page: Employ the World →
Continued on the Cooperative page: The Cue Deck →
More proofs and how it works: Proofs → · How it works → · Read the substrate → · The Substrate Record →
Works alongside ChatGPT, Claude, Gemini, or any AI you already use. All your data stays on your own computer. No cloud. Patent pending. Optionally join the Cooperative Universal Substrate™ for $5 a year. Not required.