FSL / Image Memory → Store

One half of the FSL pack's in-process image stash. FSLImageMemoryStore takes an IMAGE, saves it in memory under a string key, and passes the image straight through. Its partner FSLImageMemoryRecallSafe pulls it back out by that same key anywhere else in the graph - no wire needed between them. That's the whole point: ComfyUI graphs are rigid, and memory-keyed routing lets you break a connection without breaking the workflow.

How it works

The pack keeps a single in-process dict, _MEM, mapping key -> list of HWC tensors. On store, the node normalizes whatever it received into a flat list of [0,1] float32 HWC tensors - it untangles common shape quirks like (1,C,H,W) → HWC and grayscale (H,W) → (H,W,1) - writes them to _MEM[key], and prints a confirmation to the console with the shape and a memory id.

Two things to know about the mechanism:

• It's a pass-through. The node returns the same image on its output, so you can drop it into a chain without rerouting anything. The store is a side effect, not a fork.
• It's process memory. Everything lives in the running ComfyUI Python process. Restart ComfyUI and the stash is gone. It's not a file, not persistent, not shared across instances.

The inputs

• image - the IMAGE tensor to remember.
• key - the string it's filed under. Default "last", which is the pack's "most recent image" convention - the Nano Banana iterative workflows use "last" to loop the previous generation back in.

Output: image, identical to the input. Wire it onward as usual.

When you'd reach for it

Iterative generation is the pack's signature use: generate, store under "last", then on the next pass have the recall node feed that image back as the init image - one image building on the next without a visible cycle in the graph. It's also handy for branching: generate once, store, then recall into two different downstream branches. And it solves the "wire spaghetti" problem in big workflows where a result from stage one has to reach stage nine.

Install & caveats

Same pack, same install: ComfyUI Manager → "ComfyUI FSL Nodes", or clone into custom_nodes and pip install -r requirements.txt, then restart. No API key, no models - this node is pure local utility, and unlike the Gemini nodes it costs nothing to run.

The honest caveat is memory hygiene: the dict only gets cleaned by the pack's Clear / Clear All nodes, and stale keys silently survive run to run. If you're doing a long batch, store-then-clear at the end, or you'll find yesterday's image answering for today's recall. Also note this is not a replacement for saving files - one process restart and it's all gone.