Spotlight: Giraffe — a memory layer you have to ask for, not one that watches

Anyone who has used ChatGPT or Claude for something that spans more than one sitting knows the problem: the useful stuff is in there somewhere, buried between the 40th message and the 41st, and finding it again means either scrolling forever or re-asking the model to summarize what you already figured out. Giraffe's answer is to let you say, mid-chat, "save this to my Acme project," and have it land in a structured place: Current facts, Decisions, Open questions, Notes, Documents.

The demo on the site makes the idea concrete with two very different examples side by side. One is a family trip to Italy (skip Venice, train between cities, does the Florence hotel have a family room). The other is a B2B renewal call (their ceiling is 8%, offer 6%, open with onboarding fixes not pricing). Same four buckets, same mechanism, wildly different domains. That's a decent signal the underlying idea isn't tied to one niche use case.

The smart decision here isn't the categorization scheme itself, it's the restraint. The tagline says it plainly: nothing is saved unless you ask. Most memory tools for AI chat are trying to solve this by capturing everything automatically and inferring what matters later, which is exactly how you end up with a memory feature full of noise, half-relevant facts, and things you'd rather it forgot. Giraffe instead makes saving an explicit act, a single sentence, and treats the model as the thing that sorts it into the right bucket on the way in. It also keeps history rather than just overwriting: when a decision gets reversed, it stays on the project marked as reversed rather than vanishing, and old values of a fact are kept even after you remove the current line. That's the kind of detail you only add if you've actually hit the moment where you needed to know what you used to believe, not just what you believe now.

Exporting everything as Markdown and JSON is the other quietly good call. It means the project data isn't hostage to Giraffe's own interface. If you decide in six months you want it in Notion or in your own database, you can get it out in a form that isn't a screenshot or a copy-paste job.

Who should try this: anyone running an AI-assisted project that lasts longer than one conversation and touches more than one AI tool. The renewal-call example is a good tell. Sales reps, consultants, founders juggling several live threads of research or negotiation with an assistant, people planning something complicated (a trip, a renovation, a hiring plan) where the facts and decisions keep shifting and you don't want to re-derive them every time you open a new chat window.

Who should skip it: if your AI use is mostly one-off questions, or if you want memory to happen automatically without you doing anything, Giraffe's opt-in model will feel like friction rather than a feature. It's also worth noting that pricing, team support, and integration depth beyond ChatGPT and Claude aren't clear from the site yet, so if you need multi-seat or enterprise guarantees today, you'll want to ask before you build a workflow around it.

My honest read on where this goes next: the current pitch is strong on the individual, single-project case. The open question is what happens at scale, ten or twenty active projects, several people touching the same one, or facts that need to move between projects. If Giraffe can keep the same discipline (opt-in, versioned, exportable) while handling that kind of sprawl, it has a real shot at being the thing people reach for instead of a pile of pinned ChatGPT messages. If it stays a single-user, single-thread tool, it's a genuinely nice utility but a narrower one than the two-example homepage suggests.

---

Try Giraffe: withgiraffe.com
See the launch: Giraffe on welaunch.sh