Unlimited Context for AI
Free software that lets an AI reach a store far larger than any context window and pull back exactly what it needs. Prove it.
By address, an AI reached about 9 billion tokens of the Founder's own text (366.5 million of them counted exactly) and got the exact file 600 of 600 times, for a pointer of about 50 to 64 tokens. Condition: the address is known. Receipts PR-5 and PR-2.
Dr. Mnemosynec's Substrate Conduit, version 0.3.11 for Windows (64-bit), a zip file from our public release page on GitHub.
Have the Conduit? Join the Seven Seas: get the cards other members chose to share, computer to computer.
One Cooperative, any doorway. Membership is $5 a year. Join.
Dr. MnemosyneC
Memory specialist
The receipts
Every number below carries its condition. Each receipt lists its method, every input and output file with its SHA-256, and what it does not show.
What the pointer buys
Faster and cheaper on repeated questions. A repeated question comes back in 4.38 milliseconds instead of 35.71 seconds, more than 99% faster, and costs 97.8% less: $99.30 against $4,433.46. That is the first 300 MMLU-Pro questions asked 100 times each, against the most accurate of the flagship models from six leading AI vendors, tested on the same 300 questions (it scored 280 of 300). Condition: repeated questions only, and the cards were built for those 300. Receipts speed and cost and six flagships.
It compounds across machines. The stored work does not have to be on your own machine. Connected machines share what they have stored, so work done once on one machine answers the same question on the others. In our test, one machine pulled 20 of 20 files exactly from another machine's store, about 1 millisecond each, over a home network. In the tests that followed, every transfer ran over an encrypted, identity-checked connection, where a machine served only the files on its owner's signed consent list. The more it is used, the more it compounds. Condition: one pair of machines on one home network, with test code. Receipts PR-7, PR-8/9 and PR-10.
When it answers, it is right. When it does not know, it says so. Every answer carries its provenance, daisy-chained back to its sources, and we require the AI to say "I don't know" when it does not know. On 20 questions the Founder sealed before any search ran, it answered 12, got all 12 right and none wrong, and said "not found" on the other 8. Receipt PR-11.
THE 700, a separate blind test. On 699 more MMLU-Pro questions, each card was answered blind and hashed before the answer key was opened: 628 of 699 correct, clean-blind (89.8 percent). Every card replays with zero model calls, and you can check the hashes in your own browser. Condition: Claude models made the answers when each card was minted, scored once against the public key, and no card was edited after the key. Receipts /prove/ and the tally.
We don't fix hallucinations inside whatever AI you use, today or in the future. We make them irrelevant: every answer is stored work checked word for word against its source, or "I don't know."
| Receipt | What was tested | Result | Condition |
|---|---|---|---|
| PR-1 | Is every file on the disk in the store, byte for byte? | 400 of 400 sampled files found with the same SHA-256 | Two random samples of 200 from a fresh walk of the disk (Documents and Downloads), 2026-10-04 and 2026-10-05. A sample, not every file. |
| PR-2 | Given its address, does the exact file come back? | 600 of 600 byte-exact, for a pointer of about 50 to 64 tokens, lookup under a millisecond at the median | The address is known. Two random draws of 300 from 295,636 unique files, re-hashed from disk. Pointer tokens estimated as characters divided by 4. |
| PR-5 | How much text is reachable by address? | About 9.1 billion tokens estimated (366,528,777 counted exactly), across 286,538 unique files | By address only, with the 600 of 600 fetch result of PR-2. The 9.1 billion is an estimate (byte and character counts times token ratios measured with the cl100k tokenizer on samples); the 366.5 million is counted with the cl100k tokenizer over the prose text. |
| PR-7 | Can one machine pull exact records from another machine's store? | 20 of 20 byte-exact, median fetch 1.18 ms, index signed by the other machine (Ed25519) and verified; a copy of the index with one field changed rejected; all 6 negative controls (records that were not there, an unknown topic, a wrong access token) refused or empty | An exact topic word plus the address. One pair of machines on one home network, 20 small files, one run, on 2026-10-04 (Central). A test service, not the product's network layer. |
| PR-8/9 | Can a machine ask another for files over an encrypted, identity-checked session, and get only what was consented? | 3 of 3 impostor identities refused; 20 of 20 large files (up to 10.6 MB) byte-exact at about 17.5 MiB per second; 8 of 8 requests outside the signed consent list refused | Fetch by address over an encrypted session (X25519, Ed25519, AES-256-GCM) with test keys. One pair of machines on one home network, treated as separate networks by rule; one run on 2026-10-04 (Central); a consent set of 2,081 public files (30.3 MB). A test transport, not the product's network layer. |
| PR-11 | Twenty questions only the owner knows, about one folder, sealed before any search ran | 12 answered, 12 right, 0 wrong; 8 "not found" | 20 questions, written and marked by the Founder; every answer an exact quote from his files, checked byte for byte. 12 of 20 answered (60 percent). |
| PR-12 | Twenty hard questions about his own life and work, answered from his whole store | 19 of 20 as he meant them | 20 questions, written and ruled on by the Founder; answered by a Claude helper driving model-free search tools. Not sealed before the run began. |
| Speed and cost | A repeated question, answered from a stored card instead of a fresh model call | 4.38 ms against 35.71 s; $99.30 against $4,433.46 (97.8% less) | The first 300 MMLU-Pro questions, each asked 100 times, against the most accurate of six flagships (280 of 300). Repeated questions only; the cards were built for those 300. |
| THE 700 | 699 more MMLU-Pro questions, each card answered blind and hashed before the key was opened | 628 of 699 correct (89.8 percent), replayed with 0 model calls | Clean-blind count. Claude models made the answers when each card was minted; scored once against the public key; no card edited after the key. Kept separate from the cost number. |
These receipts, with every file hash: the receipts page.
The idea in four steps
- Application builders cannot enlarge a hosted model's context window. Most AI memory products work around that limit with retrieval (vector or keyword search), summaries, saved notes or state, longer-context models, or provider prompt caching. Ours keeps the originals instead: every record is an Eblet, a lossless copy addressed by its SHA-256 hash.
- A deterministic harness, ordinary code on your own machine with no model calls, does the record keeping: finding a record, checking its hash, applying access rules, and assembling the evidence. The model then spends its tokens on interpretation and decisions, not on managing memory.
- When the address is known, recall is exact. In our test, 600 of 600 requested records came back byte for byte, with a pointer of about 50 to 64 tokens each.
- Then nodes connect. On the night of October 4 a second machine on our network built its own crystal over 20 new files that exist nowhere else (checked: 0 matches among 14 million files on the first machine). The first machine searched that crystal and fetched all 20 records by address, byte for byte, in about 1 millisecond each over the local network, each pointer about 50 tokens. The second machine signed its index with its own key, the first machine verified it and rejected a tampered copy, and all 6 negative controls (records that were not there, an unknown topic, a wrong access token) were refused or came back empty. Only the records asked for left that machine. Then we ran it the other way, treating the second machine as a stranger on another network: the two machines opened an encrypted, mutually authenticated session (X25519 key agreement, Ed25519 identities, AES-256-GCM), refused 3 of 3 impostor identities, and the second machine fetched files from the first machine, up to 10.6 MB each, all 20 byte for byte at about 17.5 MiB per second. The first machine offered only a signed list of consented, secret-scanned public source files, refused 8 of 8 requests for anything outside that list, and sent nothing but what was asked for. Each record carries its source chain; hashes, signed append-only records and consent gates give integrity, traceability and access control, though they do not by themselves prove legal ownership or that a source is true.
Prompt caching and the substrate work together
Prompt caching is a feature of the model provider. The substrate sits on your side. We use both: fetch the exact records you need from the substrate, put them in a stable prompt prefix, and let the provider cache that prefix.
| Provider prompt caching | Substrate | |
|---|---|---|
| What is kept | The provider's processed form of a prompt prefix you send again | Lossless copies of your records, each addressed by its SHA-256 |
| Where it lives | The provider's servers | Your machine, and other machines you choose to connect |
| How long | A time limit set by the provider, often minutes to an hour | Until you delete it |
| Size | Inside one model's context window | Not tied to a window: the model reads only what is fetched (PR-5: about 9.1 billion tokens reachable by address, estimated) |
| What it saves | Cost and time of re-reading the same prefix | Finding and re-sending records: a pointer of about 50 to 64 tokens brings back the exact record (PR-2, address known) |
| Exactness | The same prefix is reused as sent | Byte-exact, checked by hash (PR-2: 600 of 600) |
Run it on your own folder
Play 20 Questions in the Substrate Conduit on your own computer. Pick a folder. Ask 20 questions only you know the answers to, and seal them before anything runs. The search answers each one from your folder: the exact passage, the file and its SHA-256 fingerprint, or "not found". You mark each answer Right, Wrong or Not found, and keep the receipt.
- Your files stay on your machine.
- See how it went for the Founder: Play 20 Questions: 12 answered, 12 right, 0 wrong and Twenty Questions, first run: 19 of 20 as he meant them.
Get Your Hands Dirty, and HELP
Sponsor the work of Liana Banyan and its 16 Initiatives for $20 a month. Choose 1 to 50 sponsorships at checkout.
A sponsorship is a gift. It is not a membership (that is $5 a year), it buys nothing back, and it is not tax-deductible: Liana Banyan Corporation is not a 501(c)(3). Payment goes to Liana Banyan Corporation through Stripe. Cancel anytime yourself: manage or cancel your sponsorship with the email you paid with.
Receipts and proofs
- Prime Radiant receipts, each with its method, every file hash and what it does not show
- 5.32 billion tokens at 50 tokens: 67 of 67 exact by address, in an earlier, separate store
- Prove It Here: THE 300 and THE 700, replayed in your browser
- Proofs and receipts and earlier receipt pages
- Sealed prediction, 2026-09-16
- The Method Deck
Step One of How to Save the World in Six Easy Steps
From the paper by Jonathan Jones, Founder of Liana Banyan, section 2, "Level the Field (Step 1)", parts 2.1, 2.2 and 2.4, quoted as written except one word: "dramatically" is left out of 2.1.
Level the Field (Step 1)
What it is
A shared memory substrate layer that makes every AI, both flagship and free, more accurate on cooperative knowledge tasks, faster, and at a fraction of the usual cost.
Why it is first
Without a shared substrate, steps 2 through 6 are coordination problems that quietly drain resources. MnemosyneC is the force multiplier that makes every subsequent step cheaper to run and easier to replicate in any locale, from San Antonio to Nairobi, on the same underlying architecture.
The rung doctrine
MnemosyneC is the first rung, not the destination. The destination is cooperation at scale, and leveraging the MnemosyneC Advantage is a huge first step.
The biggest cost in deploying AI today is not the model; it is the context. As users we rebuild state from scratch on every call, reread the same documents, reconstruct the same understanding, and pay for the same tokens again and again. MnemosyneC is the cooperative's answer to that waste.
When one AI resolves a question and deposits the answer into the substrate, every other AI that asks the same question gets the answer back for free, or close to it. The work is done once. The benefit accumulates. The substrate compounds. This is what "cooperative" means at the infrastructure layer: we do not each just build our own memory, rather we share correct answers, so questions don't have to be asked over and over.
On questions already answered once: more accurate (300 of 300 vs 280, cards built for those 300), faster (4.38 ms vs 35.71 s), a fraction of the cost (97.8% less over 100 asks).