Plain-language license FAQ. Published 2026-09-17. Not part of any license. Not legal advice.

MnemosyneC License FAQ (plain language)

This page explains the license in everyday words. It is not the license. If anything here differs from LICENSE.md, COMMERCIAL_LICENSE.md, or NETWORK_TERMS.md, those documents win.

The license in one minute

What is the license called? The Mnemosyne Cooperative Source License 2.0. SPDX identifier LicenseRef-MCSL-2.0. It covers MnemosyneC and the Liana Banyan Platform, which are one work.

Is it open source? It is source-available. You can read, run, change, and share the code. It is not an OSI open-source license, because commercial use needs a signed license.

Who uses it free? Individuals and non-profits, for anything that is not Commercial Use. That includes hobbyists, students, academic and non-commercial researchers, and open-source projects. Government agencies are not in this group (see below). A business of any kind may also use it internally for free while it and its affiliates took in USD 100,000 or less in the last twelve months. Money raised from investors, borrowed, or received as a grant counts toward that USD 100,000 just like sales, so a funded company with no sales yet is over the line once it has raised more than that. Free means no fee, ever, and the patent pledge to you is irrevocable.

Does being a cooperative or a Liana Banyan member make it free? No. What you do decides, not what you are. Any for-profit activity pays a share of savings, whether a corporation, a cooperative, a non-profit, or a person does it. The USD 5 membership is a discount on the Commercial License, never a free pass. A person running a for-profit business is doing business, so that business pays once it makes Commercial Use.

What counts as Commercial Use? Any of these, by anyone: running it as a hosted service for others; putting it inside a product or service that makes money, even a free one; using it in revenue-generating AI infrastructure; using it or its output to train, improve, or provide a commercial AI model, product, or service; any internal use, including evaluation, pilots, and research, by a business that, with its affiliates, took in more than USD 100,000 in the last twelve months, counting sales, investment, loans, and grants; connecting to the Substrate Network as a business; any use by a government agency, other than teaching and non-commercial research at a public school or university; or doing any of that for a business as its contractor, consultant, subsidiary, or front. The exact list is LICENSE.md 4.1. Splitting a business into small pieces, or routing the work through someone who is free, does not make it free.

What does a commercial user do? Sign the Cooperative Endorsement Commercial License, or do not use the software commercially. Liana Banyan Corporation grants that license under its own license from Upekrithen LLC, which owns the software and the patent applications. There is no third option. Publishing your own source code does not replace the license.

What does it cost? A commercial user pays a share of the net savings the software is verified to produce, never a share of revenue. Starter has no minimum; the committed Enterprise and Provider track adds a signing payment and a yearly minimum.

Track 1, Starter (self-serve, for companies that spend under USD 1 million a year on AI). 30 percent of your verified net savings, billed monthly. No savings, no fee. Month to month: either side can stop it with 30 days’ notice. No signing payment and no yearly minimum. Prepaying is optional: you can prepay from USD 500, credited dollar for dollar with no bonus (the rate lock below is the reward for prepaying), and the credit stays on your account against future bills, never expires, and carries over if you move to Track 2, with no cash value. The first 25 companies that prepay keep the 30 percent rate for 12 months. Your AI spend is measured from the same provider billing exports as your baseline. If it reaches USD 1 million a year, Starter continues for 60 days after we tell you, and then Enterprise terms apply: you sign a Track 2 order form or stop commercial use. You sign up for Starter online: you click “I agree”, tick a required box confirming you have authority to bind your company, and we email you the terms.

Track 2, Enterprise or Provider (3, 5, or 10 years). On the first USD 5 million of verified net savings each year: 25 percent on a 3-year term, 20 percent on 5 years, 15 percent on 10 years. Savings above USD 5 million a year: 15 percent. Above USD 25 million a year: 10 percent. The rate never goes below 10 percent. Each rate applies only to its own slice, like tax brackets, so earning one more dollar never raises the rate on the dollars before it. The 10-year term needs a prepayment of at least one year’s minimum and includes reopeners at years 3, 5, and 7 and a technology-refresh clause.

Track 2 also has a scoped pilot, a signing payment, and a guaranteed yearly minimum, set by your annual AI spend:

Each year you pay the higher of your share of savings or the minimum, never both. The signing payment is credited against your fees; the USD 2,500 activation fee, and any pilot fee above it, count against your first year’s fees. The minimum is owed for every year of the term you sign. If you leave early, you either keep paying the minimum for the rest of the term, or pay a one-time exit fee of half of the minimums left. Invoices are quarterly, after the fact, due in 30 days: each one is your share of the year’s savings so far, minus what was already invoiced that year.

Example (made-up numbers, only to show the math): a company in the USD 10M+ band with USD 1 million of verified net savings in a year pays USD 250,000 on a 3-year term, USD 200,000 on 5 years, or USD 150,000 on 10 years. With USD 10 million of savings on a 3-year term it pays USD 2,000,000 (25 percent of the first 5 million plus 15 percent of the next 5 million).

Signing fast. Speed rates apply to the first USD 5 million of verified net savings each year, and are written as exact rates on the order form. The clock starts on the day we first send you the order form for signature. The order form stays open for 30 days; if it is not signed by then it expires, and sending it again does not restart any clock. After it expires we can send you a new order form at standard rates at any time, but the speed rates and the founding-cohort offer never come back for your company.

• The first 10 companies only (the founding cohort) that sign and pay the signing payment, in cleared funds, within 10 business days: on a 3-year term, 20 percent in years 1, 2, and 3; on a 5-year term, 15 percent in years 1 to 3, then 20 percent; on a 10-year term, 10 percent in years 1 to 3, then 15 percent.
• Anyone who signs and pays within 30 days: on a 3-year term, 25, 20, and 20 percent in years 1 to 3; on a 5-year term, 20, 15, and 15 percent, then 20 percent; on a 10-year term, 15, 10, and 10 percent, then 15 percent.
• After 30 days: the standard rates. Example (made-up numbers): USD 1 million of verified net savings every year in the USD 10M+ band costs USD 1,000,000 over a 5-year term at standard rates, USD 900,000 within 30 days, or USD 850,000 for the founding cohort.

Prepaying. Every USD 1 prepaid within 10 business days buys USD 1.10 of Prepaid Fee Credit; within 30 days, USD 1.05; after that, USD 1.00. A founding-cohort company can get USD 1.15 only if it does not take a speed rate; anyone who takes a speed rate gets at most USD 1.05. Only cleared funds count. Prepaid Fee Credit cannot be refunded (except if we materially break the agreement and do not fix it within 30 days, or, for unused credit, if we end the agreement for our own convenience), has no cash value, cannot be transferred, and, on Track 2, expires at the end of your term (on Starter it never expires and carries over to Track 2). It can pay at most half of any yearly minimum; the other half is paid another way. Prepayment is capped at the lower of twice your first-year minimum or a fixed amount on your order form. The founding-cohort terms are a public offer open to anyone, for a limited time: only the first 10 companies, each within 10 business days of receiving its order form. The offer closes when 10 companies have qualified or 90 days after public launch, whichever comes first.

Every Track 2 licensee also funds Mirror roles (below). A Liana Banyan member gets 20 percent off the rate on each slice, after any speed rate, but discounts never stack below 10 percent on any slice, and never below the minimum. The terms are a draft: counsel reviews them before anyone signs.

Which track is for me? Starter if your company spends under USD 1 million a year on AI and you want to try it with no commitment and pay month to month. Enterprise or Provider if you want a lower rate in exchange for a committed term, a signing payment, and a yearly minimum. An AI company that runs the software inside its own service usually takes Track 2 as a Provider. You can move from Starter to Track 2 at any time by signing a Track 2 order form.

How are my savings measured? Verified net savings are what the same work would have cost you at your baseline, minus what it actually cost, minus the cost of the cache and of setting it up. Only costs invoiced by third parties count, and your own staff time is left out of both the savings and the costs. The baseline comes from 6 to 12 months of your own provider billing exports (on Starter with less than 6 months of history, your token volume over the first 30 days at the prices you actually pay); you get a statement every month, 30 days to dispute any statement or invoice, one audit a year at most, and an independent expert if we cannot agree. The baseline is reset only for causes that have nothing to do with the software: a volume change of more than 20 percent, a switch of model or provider, or a merger. A reset baseline is your current token volume at the prices you actually pay your providers now, including negotiated discounts, at the model mix you used before MnemosyneC; a model you no longer buy is priced at the last price you actually paid for it, adjusted by how much the provider’s list price has changed since then; our effect is left out, and a provider’s price cut lowers the baseline. Late payments carry interest of 1 percent a month or the state legal maximum, whichever is lower, as a separate line on your next invoice, and service can be suspended for nonpayment.

Is Prepaid Fee Credit the same as Liana Banyan Credits or Marks? No. Prepaid Fee Credit is a billing credit: money a company pays us early, applied to its own license bills. It is not a member Credit, a Mark, or a Joule, and it can never become one. Marks, Credits, and Joules are never converted to fiat money. Ever. Marks can only be spent in the platform or given to another member; trading a Mark for money or anything else of value, anywhere, is not allowed and voids that Mark. Anyone can earn Marks; spending them requires the USD 5 membership.

Who pays: the AI company, or its customers? Whoever runs the software pays, once. An AI company that runs it inside its own service holds a Provider License and pays its share of its own measured savings. Every customer of that service is then covered for using it, businesses and federal agencies included, with no separate license and no extra fee. If you run the software yourself, on your own hardware, node, mesh, or install, that is a Direct Deployment: you pay your share of your own savings and, on Track 2, fund your own Mirror roles. Using a licensed provider does not cover a deployment you run yourself. No dollar of savings is charged twice.

Does the AI company get off cheap by passing the savings on to its customers? No. A provider’s savings are valued at the price of the work it no longer has to do, its list or contract price, not at its own cost. If it passes those savings on as lower bills, they are still counted once, at full value.

What is the Mirror Clause? For every role you fill to install and run the software, you pay for one role held by a Liana Banyan cooperative member, for the same period, at no less than that role’s pay. You pay it whether or not you give the member install work. The salary always buys real work: the member does your install work, or other work you need, or, when you have none, work for the cooperative’s Initiatives through Employ the World. The pay goes to the member through Employ the World. Liana Banyan never holds it. It is a requirement of every Track 2 (Enterprise or Provider) Commercial License and of the Government Mirror Program, not a target. It does not apply to Starter.

Do government agencies pay? Yes. Government agencies are not non-commercial. An agency that saves money pays like a company. The Federal Government, and every other country’s national government, pays its share of its savings on any deployment it runs itself, like any licensee, and nothing extra when it uses a licensed provider’s service. Other countries’ national governments must also pass US sanctions and export-control screening.

Can a state, county, city, tribal, or territorial government, or the District of Columbia, use it without the fee? Yes, through the Government Mirror Program, for as long as it does three things: follows the Mirror Clause; adopts the cooperative’s sixteen Initiatives as the way it delivers those services where they fit its work (The Kitchen Table, for household meal planning, is one example); and hires cooperative members, under contracts worth at least what the fee would have been (its share of verified net savings at the standard rates, and at least its yearly minimum on Track 2), the way the Army hires civilian firms. The Harper Guild decides who is hired and whether the rules are kept: every job is posted openly to all members, the government cannot pick the people, and anyone with a conflict of interest has to disclose it and is left out. No cronyism. If a government falls below the line, it gets written notice and ninety days to fix it; if it does not, the fee starts from then on, never backdated. The Federal Government cannot join this program. State, provincial, and local governments of other countries can join this free program, on the same conditions as tribal and territorial governments, only if their country is not under comprehensive US sanctions, they are not on the US sanctions list or the Commerce Entity List, and they sign an export-compliance statement. Tribal and territorial governments and the District of Columbia get a minimal license: no indemnities, nothing they cannot legally agree to, and no procurement contract needed for a free license. They still follow the Mirror Clause and the Initiatives, still hire cooperative members under contracts worth at least what the fee would have been, and the Harper Guild still decides whether the rules are kept.

How long does it last? Starter runs month to month until either side gives 30 days’ notice. Enterprise and Provider terms are 3, 5, or 10 years. Your rates and the measurement method are locked for the term you choose; on a 10-year term either side can reopen the measurement at years 3, 5, and 7 (a reopener reopens only the measurement; the rates change only if both sides agree in writing, and never below 10 percent), and a technology-refresh clause resets the measurement (never the rates) if your providers’ list prices for your main models fall more than half below your baseline. After the term, the license renews a year at a time at our published rates at renewal (founding-cohort and other speed rates do not carry over), unless either side gives 60 days’ notice. If we cannot agree on a baseline adjustment for a cause that has nothing to do with the software, an independent accountant decides it for the periods ahead only; your rate never changes that way. You pay for every period you use it. If the savings stop being worth it, you stop using it and certify that you stopped; on Track 2 the yearly minimum is still owed for the rest of the term you signed, unless you end the agreement because we materially broke it and did not fix it within 30 days. When Liana Banyan Corporation’s charter ends, its licenses pass to the daughter companies it formed, under the same agreement with Upekrithen LLC, which owns the software and patents.

What is Tier 2? An optional election, offered only in the Commercial License (not in the public license), to pay an additional thirty percent of measured savings straight to a public sovereign fund set up by government, on top of the license fee. The license fee is what we require; the thirty percent is what we propose an AI company give back to the public. Liana Banyan never touches that money: it goes from the company directly to the fund, we only verify the measurement, and no Liana Banyan member receives any of it. Tier 2 is not offered until such a fund exists and is designated.

Are the patents granted? No. The portfolio is patent-pending. Until patents issue, the commercial license is enforced through copyright, and connection to the Substrate Network is governed by its own terms. Commercial fees pay for the software, its updates, service, and the savings meter, not for the patents; any part of a fee tied to a patent would drop away automatically if that patent expires or never issues.

Can someone rebuild the ideas without the code? Copyright protects the code we wrote, not the ideas in it. Someone who writes their own software from scratch, without copying our code, without connecting to the Substrate Network, and without signing anything with us, is outside this license. What reaches them is the patent portfolio once patents issue, and nothing before that. We do not claim otherwise.

I downloaded an earlier version. What applies to me? Versions before 0.9.31 were released under the earlier license (version 1.0). Those copies keep the terms they came with. Version 2.0 applies from 0.9.31 forward, and the Substrate Network Terms apply to every connection from their effective date, whatever version you run.

Where do I ask? licensing@mnemosynec.ai for the Commercial License. legal@mnemosynec.ai for trademark and enforcement questions.

Performance: what is measured, and what is not

The license makes no performance promises. Earlier license text (version 1.0, section 6.2) listed performance figures inside the license. Those figures are removed. This section says what has actually been measured, with its limits. Full method and receipts: mnemosynec.org/cost-collapse.

Figures from the old license that are withdrawn

These appeared in version 1.0 section 6.2. None has a current receipt on the measured list, so none is claimed:

• 30 to 60 times faster wall-clock processing
• about 100 times lower token cost per query
• 90 to 98 percent fewer data-center round trips
• 0.059 ms mean query latency
• about 26,000 times context-window amplification
• 100 percent disaster-recovery byte-perfect restore
• “low-cost equals high-end” cross-model accessibility

No general cost-collapse ratio has been measured. We do not state one. Every figure below is one measurement on one date with one method; results vary by workload, deployment, baseline assumptions, hardware, model, and operating conditions.

What is measured (2026-09-16 unless noted)

1. Stored answers run as code. Each of our first 300 real MMLU-Pro questions has a stored card that code runs. All 300 scored correct against the answer key, 284 blind-labeled and 16 KEY_INFORMED (repaired after a blind miss with the key in view, and labeled), with 0 model calls. Limits: this is recall of questions already carded; one card fits one question and nothing here measures a new question. The first 150 cards were reworked while scored against the key, so their blind label is weak. Many cards are table lookups on a reading stored before the key was read.

2. Provenance, two separate numbers. All 300 answers are hash-verified back to their stored card and replay byte-for-byte on two machines with no model call. Limits: only 4 of 300 carry a signature a reader can re-check today. Only 25 of 300 cards cite a source a stranger can check (a URL, a DOI, or an edition with page or section); 225 cite only our own reading. We do not claim “100 percent provenance”.

3. Addressed recall does not grow with corpus size. Given its address, the system retrieved 67 of 67 real records from a 172-million-token corpus (164 times the largest flagship context window we use), at a flat 50 tokens per question from 1,000 to 100,000 files. Limits: finding a record without its address is still a search problem, and that search got worse as the corpus grew (top result right 49, then 46, then 27 of 67). No model answered anything in this test; it measures delivery and retrieval, not reasoning. We do not call this “unlimited context”.

4. Method cards help flagship models, on a small set. On the 19 of 20 hard questions every vendor answered, all six flagships scored as high or higher when given our method card (+1 to +9 of 19). Anthropic declined to answer the 20th with the card in view. Limits: n is 19; the questions were chosen because flagships missed them, so some gain would appear with no card at all; there was no same-day plain control; 2 of the cards are KEY_INFORMED. It is a small validation, not a rate.

5. Two machines share stored answers. One machine copied the 300 cards and the code that runs them to a second machine over the local network, signed and hash-checked (310 files sent, 0 refused). The second machine produced the same 300 answers, byte for byte, with no model called. Unsigned, unenrolled, impersonated, stale, or altered requests were refused. Limits: two machines, one owner, one LAN, nothing over the internet. Throughput and cost effects of the mesh are not measured.

6. One warm-versus-cold pair (2026-08-20). Same question, same local model, same machine: total wall clock fell from 275.9 s to 79.7 s (down 71.1 percent), live search calls from 16 to 0. Limits: n is 1. The speedup is in the search leg only; the answer leg was 9.4 percent slower warm. Metered dollars were $0.00 on both runs for architectural reasons (a signed-in browser and a local model), with electricity and hardware excluded. The answer was wrong on both runs; this was a cost receipt, not an accuracy receipt.

What your commercial fee is based on

None of the figures above. A commercial fee is based only on your own measured savings, from your own receipts, under Exhibit A of the Commercial License.