Your AI has Amnesia. Dr. MnemosyneC has the Cure.
Mnemosyne (neh-MOZ-uh-nee) is the Greek goddess of memory.
Every time you start a new session, your AI forgets everything. Your projects, your preferences, your past conversations, gone. Dr. MnemosyneC gives your AI a permanent, private memory that actually stays.
- Works alongside ChatGPT, Claude, Gemini, or any AI you already use.
- All your data stays on your own computer. No cloud. No account required.
- Free to use. A $5/year membership unlocks the full cooperative.
Free forever · No ads · No strings · Data stays on your computer
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.
Does it actually work?
PROVE IT YOURSELF75 questions · 4 AI vendors · real test run 2026-05-30. No tricks.
The Banyan Metric measures recall of substrate-stored information. "Without" = model alone, cold start. "With" = same model after MnemosyneC substrate context injected. The substrate does the knowledge-recall work; the model reads and answers.
The free local model (Ollama, $0 inference cost) lifted from 6% to 78% entirely because of the shared substrate. Full proofs →
📌 Pinned proofs
Plain receipts. Caveats stated aloud. No hype.
← scroll for more · All proofs →
Download MnemosyneC v0.1.60 — free, forever
▼ Download now
Starter version
456 MB · installs quickly, works immediately. A small built-in AI model is included so you can use it right away with no extra setup.
qwen2.5:0.5b bundled
In-app upgrade
Full version
Upgrade to the full version inside the app anytime, free. Bigger download, noticeably better results, same privacy guarantee.
Gemma 4 12B via Ollama (~7 GB pulled in-app)
⚠ Wanted
Mac build of MnemosyneC
Build targets are configured (dmg + zip, x64/arm64). Be the builder who ships it. Earn Marks.
View Mac bounty →⚠ Wanted
Linux build of MnemosyneC
deb/rpm/AppImage targets ready. Be the builder who ships it. Earn Marks.
View Linux bounty →🎬 Wanted
YouTube video: how to use MnemosyneC
Record a clear walkthrough: install, first launch, offline query, mesh connection. Show real results. Post publicly. Earn Marks and Founding-Contributor credit.
Claim this bounty →Windows ships now. Mac and Linux community builds coming.
One quick thing before you run the installer
When Windows sees new software for the first time, it shows a caution screen. This happens with all new programs, including ours. It does not mean there is a problem. When you see it, click More info and then Run anyway to continue. The installer is safe.
- Click the download button above and save the file. (Large file: the bundled AI model is included.)
- Open the downloaded file. When the Windows caution screen appears, click More info.
- Click Run anyway.
- Follow the installer prompts.
- Done. Find MnemosyneC in your Start menu.
Version 0.1.60 · Free forever · No account required · All data stays on your computer
▶ Technical methodology — SHA‑256, Gauntlet, patents, substrate, licensing
Binary integrity · v0.1.60 · SHA‑256 checksums and signing
| Field | Starter (qwen2.5:0.5b bundled) | Full (in-app upgrade) |
|---|---|---|
| Filename | MnemosyneC-Setup-0.1.60.exe | No separate file — upgrade in-app via Settings |
| File size | 456 MB | ~7 GB pulled via Ollama in-app |
| SHA‑256 | See GitHub release page for verified hash | N/A — verified by Ollama registry |
| Signing | DigiCert RFC 3161 · .tsr sidecar — see Release Notes | N/A |
| Hosting | GitHub Releases (v0.1.60) | Ollama registry (pulled in-app) |
| Telemetry | None without explicit opt‑in | |
Verify via PowerShell:
Get-FileHash "MnemosyneC-Setup-0.1.60.exe" -Algorithm SHA256Benchmark methodology (BP063 R10 Harness)
75 factual questions about a single individual (the Liana Banyan Founder) were run through 4 vendors (Anthropic, OpenAI, Ollama local, Google) in both cold (no substrate) and hot (with MnemosyneC substrate) conditions. All questions test recall of manually-ingested substrate-stored information, not general knowledge or training data. Inter-rater agreement between the primary grader and a spot-check grader: 0.865–1.000 (strong to perfect, Cohen's kappa). The full reproducibility pack is at github.com/liana-banyan/lb-reproducibility-pack.
USPTO patents · Cooperative Defensive Patent Pledge #2260
MnemosyneC is covered by 21 USPTO Provisional Patent filings (sole inventor, Conductor-Class doctrine, Thaler v. Vidal). The full patent portfolio is pledged under Cooperative Defensive Patent Pledge #2260. What that means: cooperatives can use, improve, and deploy MnemosyneC without fear of enclosure. The pledge is irrevocable. No future board, acquirer, or successor entity can weaponize these patents against cooperative-class users.
SSPL v1 license
Run it. Modify it. Self-host it. If you offer MnemosyneC as a networked service to others, you must share your modifications under SSPL v1. This is the cooperative commons lock. It prevents enclosure from re-entering through the side door. View source on GitHub →
The Gauntlet · six-stage verification
The Caithedral™ Inspector inside MnemosyneC's Developer Tab runs a six-stage testing framework verifying substrate integrity across every build. You can run it yourself, without an LLM, to verify the substrate functions correctly on your machine.
Substrate architecture (brief)
MnemosyneC ships with a 16-component cooperative substrate including: Eblet™ ROM-class permanence layer, Wrasse sub-millisecond pre-injection (0.059ms mean), Pheromone event-class signaling, Caithedral™ Inspector, Hearth (on-device Ollama fallback), Drekaskip wave generator. Full detail: Banyan Almanac Issue 005 →
Download on GitHub
View Release v0.1.60 on GitHub → · Release changelog → · K533 Reproducibility Pack →
► The numbers behind the chart
The benchmark is real data, not marketing copy. BP063 R10 Harness ran 75 factual questions through 4 vendors in both cold (no substrate) and hot (with MnemosyneC substrate) conditions. Inter-rater agreement between the primary grader and a spot-check grader: 0.865–1.000 (strong to perfect). The delta is a real signal, not grader noise.
In addition, the Cooperative Architecture benchmarks show 30–60x wall-clock speedup and approximately 100x token-cost reduction on real workloads (Founder-run, timestamped git commit chains, session logs). Every figure has a traceable receipt.
- Proofs and empirical receipts →
- Technical Prove It test (component-by-component) →
- K533 Reproducibility Pack (run it yourself) →
Truth-Always discipline: all empirical claims are sourced to timestamped session receipts. Numbers are never estimated; they are measured or they do not appear.
► Designed to be copied and improved
A defensive moat that gets stronger the more we file
Every provisional patent filed is priority-dated prior art, a date-stamped stake in the ground that prevents anyone from coming in afterward and fencing the commons. The more we file, the wider the protected territory. The moat does not protect us from users; it protects users from enclosure.
The whole architecture is built to be copied and forked: SSPL + Pledge #2260. Run it. Self-host it. Improve it. The license is designed to let you build on it without fear and to prevent anyone from quietly re-privatizing the substrate underneath you.
We patent so no one can fence it, then give it away so everyone can build on it, deterrence and invitation in one.
One ledger. One path.
There is one ledger and one path. Every filing enters the commons and stays there.
The one way to sponsor a patent into the commons
There is exactly one way to "acquire" a patent: sponsor it for $5,000 through Upekrithen, LLC, structured as an irrevocable donation into the commons, Founding-Sponsor recognition plus Defensive Patent Pledge #2260. This is not ownership and not extraction rights.
You can fund a patent into the commons. You cannot buy one out. Every sponsorship makes the commons stronger, never smaller.
Truth-Always: we do not say "never for sale." We say "never sold for extraction." The one acquisition path is a donation into the commons, not extraction out. The distinction is the whole point.
► What does AI actually cost?
Three dimensions most people never see at once.
1 · What it costs you — savings made visible
In the benchmark, a free local model answered 78 out of 100 questions correctly at $0.00 because the MnemosyneC substrate did the knowledge-recall work.
Think of the barrel as your AI query budget. Every query answered from the cooperative substrate stays full for free. Only queries that go past the substrate cost anything.
Your current AI subscriptions hide their cost from you. You pay a flat fee and never see which queries burned expensive compute vs. which ones could have been answered from memory. MnemosyneC makes the savings visible, daily.
2 · What it costs to develop — the Substrace Theorem
Building with AI today means re-explaining everything from scratch every session. The AI forgets. You pay to remind it. You pay again next session, and the session after that. This is the hidden subscription tax: not the monthly fee, but the cost of the forgetting.
The Substrace Theorem states: when a shared substrate holds the context, any model at any endpoint, even a free local one, can reconstruct the same state from the same inputs. The cost of re-derivation approaches zero. MnemosyneC is the implementation of that theorem.
3 · What it costs in resources — DECENT, no new data centers
Every query routed through the cooperative substrate instead of a centralized cloud model is a query that doesn’t require a new GPU farm. The cooperative substrate runs on your own computer, hardware you already own. When enough members share a substrate, the mesh of existing machines is the data center.
This is the 2nd Second Industrial Revolution framing: decentralize intelligence the same way 3D printers decentralized manufacturing. Your own librarian. Your own fab shop. No new buildings required.
► The 2nd Second Industrial Revolution
"The First and Second Industrial Revolutions centralized production: big factories, one-size-fits-all. The 2nd Second Industrial Revolution de-centralizes it: 3D printers and local tools let any town make its own things. Your AI librarian does the same for knowledge. Your own library. Your own tool. Run it yourself. Every town its own fab shop, and its own librarian."
MnemosyneC is a node in that revolution. Every install is a library that belongs to its owner. Every member who contributes to the shared substrate is depositing knowledge into a commons that lifts everyone's accuracy, regardless of model vendor, regardless of subscription tier.
The platform keeps cost plus 20%. Workers, Builders, and Creators keep 83.3% of every transaction. Membership: $5/year. No venture capital. No ads. 50-year sunset clause: at year 50, the corporation dissolves and the cooperative commons inherits everything.
Why join the cooperative →► The Library of Congress Project
The Library of Congress holds records of approximately 49 million items: 17 million books and other print materials, 12 million photographs, 5 million maps, 3 million recordings. Most of it sits behind professional barriers, access requires a researcher's credential, a library card, a subscription, or a plane ticket to Washington, D.C.
The Library of Congress Project is Liana Banyan's first Grand Project: Eblet the public domain LOC catalog and make it freely searchable by any member through the cooperative substrate, in record time, at record low cost, with record accuracy. No new buildings. No new data centers. No extraction. Proof to the People.
The network target is 10,000 working nodes. The founding circle is open now.
Learn more about the LOC Project →Bounty posters: open slots
Founder fills rewards. Community fills the work. Earn Marks for making it happen.
🍂 Wanted: YouTube tutorial
"How to Use MnemosyneC" — a quality walkthrough showing install to first query. Must show: install, first launch, offline query, and mesh connection. Published to YouTube.
Earn: Marks for a quality tutorial · Apply: bounties page
🍎 Wanted: macOS port and signing
We have no signed Mac build. A macOS release requires: Apple Developer ID, Mac hardware, willingness to maintain the build pipeline. This is real infrastructure work, compensated with Marks.
Earn: Marks (significant) · Apply: bounties page · No macOS download button will appear until a signed, notarized build exists.
🐧 Wanted: Linux AppImage tester
Help validate a Linux AppImage build of MnemosyneC. Requires: a Linux machine (Ubuntu 22.04+ or Fedora 38+ preferred), willingness to run install-to-first-query flow and report results.
Earn: Marks for verified test reports · Apply: bounties page
► Glossary: terms you may not recognize
A quick reference for the vocabulary used here that you will not find in a standard dictionary.
| Term | What it means |
|---|---|
| Eblet | A single immutable record in the cooperative substrate. Every fact, file, or event is an Eblet, sha256-stamped and append-only. |
| Socceri | Your member address inside the cooperative. Your identity on the network, portable and self-sovereign. |
| Yoke | The structured message-passing protocol between cooperative AI agents. A Yoke Return is a formal reply that closes a task loop. |
| Coffee | An informal cooperative gathering, in-person or remote, where members connect and build relationships. The smallest unit of community. |
| Banyan Metric | The composite score for cooperative AI quality: accuracy, trust-calibration, and cooperative-class behavior measured together. |
| Caithedral Effect | The phenomenon where each member's substrate contribution improves every other member's AI accuracy. The cooperative knowledge commons compounds. |
| Marks | The cooperative's participation record. Marks track work performed on unfunded projects and serve as contingent accounts payable, not equity or investment shares. |
| Substrace | The theorem stating that when a shared substrate holds context, any model at any endpoint can reconstruct the same state from the same inputs. The cost of re-derivation approaches zero. |
► Who made this, and why it is not a faceless corporation
There is no team of forty. There is no go-to-market deck, no Series B, and no venture capital waiting in the wings. There is one person who has spent the last two decades circling a single question: why does technology always end up extracting from the very people who use it?
In his own words: "I can't do everything, but I can do something. And this is what I'm doing." He built the tools he needed, only to realize the framework underneath them was far more valuable than the applications on top, so he documented it all and filed twenty-one provisional patents. Then he made a decision that breaks every rule of modern tech. Rather than sell it to the highest bidder, he is giving it to the users.
The legal entity, Liana Banyan Corporation, is built with a fifty-year sunset clause hardcoded into its bylaws. At year fifty, the corporation dissolves. "The Liana," the cooperative commons, the technological substrate, the patents, and the member network, inherits everything. This isn't a marketing promise; it is a structural commitment.
The ethos is simple: be who you needed. Build the bridge behind you. Half a century from now, the corporation will dissipate into the cooperative we are growing today. The founder doesn't want to be famous. He just wants the engine to work, for all mankind, exactly as it was designed to from the start.
Not left or right. A more effective team.