Image memory (Theia)

Theia is Mnemosyne's visual memory. Once it is on, the images in your watched folders are indexed, and a question like "show me the diagrams that look like this" brings thumbnails back under the answer.

It is off by default. Turning it on happens entirely in Settings then Images: the switch installs the engine, downloads the weights, then starts indexing. While it is off, nothing else changes in how the app behaves.

The model does not see the images

That is the limit to keep in mind. Theia finds images that match your request and shows them to you, but the model writing the answer does not see their pixels. It knows their file names, not their content. An answer claiming to describe what it sees in a thumbnail would be an invention.

The states, and what they mean

Inside a watched folder, the screen tells you what will happen to the images there: they will be analysed if the vision engine is ready, ignored if image memory is off, and left waiting if image memory is on but the engine is not installed yet.

Those three situations are different and are not fixed the same way, which is why they are not merged into one message.

Going faster

Indexing runs on the processor by default. If your machine has an NVIDIA card, Settings then Images offers to switch to GPU, which is several times faster. The switch downloads a CUDA runtime of about 2.4 GB once, unless it is already there because another engine paid for it.

When a lot of images are waiting, a card offers the switch and states the estimated time left on the processor.

Cutting out a background

Right-clicking an image on the canvas offers to remove its background. The computation is local, and the original is never modified: "Restore background" returns to the starting image in one click. Two quality levels exist, set in Settings then Images; the heavier one changes nothing on a small image and takes far longer, so small images automatically go through the light model.