Mnemosyne has its first compute sponsor. Atlas Cloud is providing $150/month in inference API credits, starting July 27th, 2026.
I want to explain this properly rather than just posting a logo, because a logo tells you nothing and I think you deserve to know exactly what changed hands.
The problem this solves
Mnemosyne makes claims about recall quality. Claims like that are only worth something if you re-measure them, continuously, against real models — because a recall pipeline can regress in ways no unit test catches. A weight change that looks harmless can quietly cost you ten points of precision, and you find out from a user, months later, if you find out at all.
The honest situation: running that suite nightly against multiple providers costs real money, and for most of this year I did not have real money. So the benchmarks ran when I could afford them, which is another way of saying they were not a safety net at all.
These credits fix precisely that. They pay for nightly recall benchmarks, multi-model parity tests — the same query set across different models and embedding backends, to catch the places where we quietly depend on one vendor's behavior — and OpenAI-compatible provider coverage in the Hermes integration.
Who they are
Atlas Cloud runs a full-modal inference platform: one API in front of 300+ curated LLM, image, and video models, so you connect once instead of maintaining a separate vendor integration per modality. That is, bluntly, a convenient thing to be sponsored by when your parity suite needs to hit a dozen models on a schedule — and it will matter more once the multimodal work lands, since describing an image or a video frame means calling something that can see.
They also run a coding plan , if you want to try the thing that is paying for our test runs.
What this does not buy
This is the part I care most about getting right, so let me be blunt.
- No default. Mnemosyne does not ship configured to route through Atlas Cloud, and it never will as a condition of sponsorship. Your memory stays local, your provider stays your choice.
- No editorial control. They do not review posts, benchmarks, or docs before publication. This post was not sent to them for approval.
- No benchmark favoritism. If a parity test shows their models doing worse on our recall suite, that result gets published exactly like any other. A benchmark you are allowed to lose is the only kind worth running.
- No roadmap influence. Features are decided by maintainers and issues, per MAINTAINERS.md.
The full terms — what we accept, what we refuse regardless of how the deal is structured, and how disclosure works — live on the partners page . This is a disclosed material connection under the FTC Endorsement Guides (16 CFR Part 255), and it is disclosed there, in the repo, and here.
Why accept sponsorship at all
Because the alternative was worse. An unfunded open-source project has three options: burn the maintainer's own money, quietly stop doing the expensive-but-important work, or take help with the terms written down in public. The second option is the one that actually happens most often, and it is invisible until something breaks.
So: thank you, Atlas Cloud. The nightly suite is running again, and every regression it catches before a release is one you never have to find in production.
If your company runs an OSS credits program — compute, hosting, observability, developer tooling — the partners page explains what we accept and what we don't. The engine stays MIT, and free, forever.