What is inference (and why it's the only part you actually touch)

Inference is what’s happening every time you send Claude a message and get a response back — the model reading your prompt and generating an answer. It’s worth knowing the word mainly to separate it from training, which is the earlier, one-time process of building the model in the first place.

As a founder, you never touch training. You only ever interact with inference — every chat, every Claude Code session, every Cowork task is one. That’s also why a memory file or a piece of context you paste in doesn’t “teach” the model anything permanent; it just shapes that one inference, that one response. Nothing you write in PRODUCT.md changes Claude itself — it changes what Claude reads before answering you, every time.

That distinction explains why memory files have to exist as actual files Claude reads, rather than something you mention once and expect it to remember forever. There’s no “remembering” between separate conversations unless you give it something to read again.