Hey again,
So I was deep in a Claude Project, working out a page layout for my Shopify store. After a long back-and-forth, we landed on a direction. Spacing, hierarchy, the kind of call that takes an hour to settle on. I went to bed planning to pick it up in the morning.
Next day, new chat, same project. Claude had no idea what I was talking about.
I ended up scrolling back through the previous night's conversation manually, hunting for the exact exchange where we'd landed on it, just to hand Claude back its own decision.
Memory Is a summary, not a transcript
Claude Projects have their own memory space. What I didn't understand is how that memory actually gets written.
Claude runs a background process that reads through your conversations and writes a summary of what it decided mattered. That summary, not the conversation itself, is what loads into your next chat. For a long time, that synthesis ran on a rough 24h cycle. Start a second chat before that cycle catches up, and you're talking in a conversation that hasn't caught up yet.
It's also selective. The digest isn't a transcript. It's Claude's summary of what seemed worth keeping, which means specifics can get smoothed over or dropped entirely.
It makes sense once you think about why memory works this way at all. Reading and summarising every conversation in real time, for every user, is expensive. A background cycle is the tradeoff that makes memory possible at scale.
What to actually do about it
Some of this changed recently. You're no longer stuck waiting for the daily cycle. Claude can now update its memory the moment you tell it to, mid-conversation, no need to leave the chat.
That fixes the lag. It doesn't fix the lossy part. A manual update still produces a summary, not a transcript.
So now:
Anything precision-dependent, exact copy, a formula, a spec, goes into a markdown file saved to Project Knowledge instead of memory. That's retrieved as exact stored text, not paraphrased.
If it's worth keeping but doesn't need to be exact, I just tell Claude to remember it, so it gets picked up by the memory system, instead of storing that detail on a md file.
Memory doesn't update instantly. A new chat opened minutes later won't have caught up yet, manual trigger or not. So I either wait a bit, or hand it the context myself.
Projects scope what Claude has context on. They don't guarantee it holds onto every detail inside that scope. I’d check in on that occasionally.
For the layout call specifically: I'd tell Claude to remember the direction the moment we landed on it. The exact spacing and hierarchy would still go in a md file. Don’t count on memory alone to hold that level of detail.
Memory tells Claude what mattered. A file tells Claude what happened. Know which one you're relying on before you need it.
Bonus Tip
The manual trigger only helps if you check it actually landed. Settings > Memory, shows exactly what's stored, grouped by category, editable or deletable entry by entry.
Ask Claude directly what it remembers about a Project whenever you're unsure. It'll tell you. After telling it to remember something mid-chat, glance at Settings > Memory to confirm the entry actually shows up.
Catch you next issue,
Hasen
Infinite Output — Practical AI for solo founders.
