ChatGPT Projects vs Claude Projects vs Gemini Gems: Which Holds Context Best

If you’ve moved any real work into an LLM workspace, you’ve hit the wall: you uploaded the spec, pasted the constraints, wrote three messages of clarifications, and then the next session starts cold. Or worse — it starts with stale context from a thread you abandoned two weeks ago.

We spent two weeks running the same six tasks through ChatGPT Projects, Claude Projects, and Gemini Gems to figure out which one actually remembers what you told it. Not in marketing copy — in usage. The differences are larger than the feature lists suggest.

How each platform actually stores context

The three products use the words “project” and “gem” loosely. Underneath, the mechanics are very different.

Claude Projects attach a knowledge base (text files, PDFs, code) to a project workspace. Anthropic exposes roughly 200K tokens of project knowledge that gets prepended to every new conversation inside that project. Custom instructions sit alongside it. Each new chat is a fresh conversation — Claude does not pull in your previous chats within the project. The shared layer is the knowledge base; the chats are independent runs against it.

ChatGPT Projects work the other way. The knowledge base feature exists (file uploads per project) but the more interesting layer is memory: ChatGPT writes facts about your project into a memory store that persists across new chats. When you start a new conversation inside a project, the assistant has read summaries from earlier threads. The downside is that you don’t fully control what gets written — it’s a model decision, and it sometimes captures things you’d rather it forget.

Gemini Gems are closer to “custom GPT” than to either of the above. A Gem is a configured assistant: system instructions, optionally a set of attached files (for Gemini Advanced), and a name. Conversations inside a Gem don’t share memory with each other by default. The compensating advantage is Gemini’s 2M-token context window — you can stuff an entire codebase into a single chat and stay coherent for hours.

What we observed across two weeks of testing

We ran the same six tasks through each platform: drafting a refactor plan for a TypeScript repo, summarizing a 70-page PDF, answering follow-up questions a week later, debugging a regression after a context reset, writing release notes from a changelog, and producing a competitive analysis from scraped pages.

Claude held instructions best. The knowledge base loaded reliably on every new chat — when we put a style guide in the project, every new conversation respected it without prompting. New chats lost the conversational context (what the previous chat decided), but the foundational context never drifted.

ChatGPT held conversational continuity best. Asking “remember how we decided to handle nulls last Tuesday?” — only ChatGPT could answer correctly without us re-pasting the decision. The memory layer caught most of the decisions we cared about. But it also captured things we didn’t want: assumptions from one debugging session bled into the refactor task two days later.

Gemini won on raw context size. When we dumped a 600KB markdown bundle into a Gem chat and asked specific questions, Gemini answered cleanly without re-summarizing. Claude and ChatGPT both started compressing or asking us to narrow scope around 150K tokens. But the moment we started a new conversation in the same Gem, all that loaded context was gone — we had to re-attach files and re-explain.

Practical translation: Gemini is a long-context vehicle, not a long-memory vehicle. The 2M-token window is real and useful, but treating a Gem like a persistent workspace will frustrate you.

When each one wins

Pick Claude Projects when you want a stable foundation across many short conversations. Engineering teams using a project to enforce coding standards, writers maintaining a voice guide, anyone who wants the same baseline applied to every chat — Claude’s prepended knowledge base is the right shape for this. The 200K-token ceiling is generous for documents but tight for whole codebases.

Pick ChatGPT Projects when the work is iterative and your decisions matter more than your documents. Product strategy, ongoing client work, research projects where you build up shared understanding over weeks — ChatGPT’s memory layer is the only one of the three that genuinely remembers what you decided. Audit the memory periodically (the UI lets you view and delete entries) so stale assumptions don’t pollute later sessions.

Pick Gemini Gems when you have one large body of material to interrogate in a single session. Reviewing a long contract, analyzing a quarter of meeting transcripts, working through a research paper with all its citations attached — the 2M-token window is unmatched. Don’t expect the Gem to remember the session next week; treat each chat as a standalone analysis.

What none of them do yet

All three platforms still treat each chat as a separate event. None of them give you a true cross-session graph of “what we’ve discussed about topic X” the way a properly indexed knowledge tool would. If your work depends on retrieving prior reasoning across dozens of threads, you’ll outgrow these workspaces — and likely end up exporting transcripts into a separate system.

The other gap: there’s no clean way to share a Project or Gem across a team yet. Claude’s team plan allows shared projects but the controls are coarse, ChatGPT Projects are personal, and Gems are tied to a single Google account. For solo work the trade-offs are clear; for teams, you’re still stitching it together yourself.