I stopped hitting Claude's rate limits after turning my Obsidian vault into its cache

Published Sep 9, 2026, 10:30 AM EDT Nolen began their writing career in 2019, with three years dedicated to editing the Creative section at MakeUseOf. Their expertise lies at the crossroads of technology and creativity, covering areas like photography, video editing, and graphic design. Outside of work, you'll often find Nolen diving into a good book, writing their own stories, or playing video games. Claude is one of the fastest ways to get things done in a folder full of files, no matter the project or work you do, but the rate limits can be pretty unforgiving. The 5-hour rolling window is a whole thing on its own, and then sitting on top of that is the weekly cap that burns through faster than the message count would suggest, because tokens don't care how many messages you sent - they care about how much context was in each one. So the question becomes what actually loads into that context, and how to keep it lean without kneecapping what Claude Code can do. Turns out, it's pretty simple. It's simply a couple of text files in my Obsidian vault that Claude Code reads before it gets to work. Want to stay in the loop with the latest in AI? The XDA AI Insider newsletter drops weekly with deep dives, tool recommendations, and hands-on coverage you won't find anywhere else on the site. Subscribe by modifying your newsletter preferences! My Obsidian vault does the remembering So Claude doesn't have to I've written a lot about using Claude inside my Obsidian vault, and it's safe to say at this point that Obsidian has stopped being just my notes app - it's now also my cache. When I say "cache" here, I don't mean a literal /cache/ folder. It's taking the concept of a cache, meaning a fast and temporary memory, and applying it to my workflow. And the way it works is literally however you want; it depends entirely on your projects, tools, workflow, and brain. For me, my "cache" is a couple of files in the vault telling Claude how to act and giving it the context it needs, and that data gets updated regularly. The setup grew piece by piece. One of my vaults has an /index/ folder Claude Code auto-generates and refreshes on command - active projects, stalled pitches, this-week, and by-area. My reading log sits in its own folder with a CLAUDE.md schema file at the root and a small SKILL.md that tells Claude Code exactly which notes to pull for which type of question. Those are just a couple of examples. And it's those pointer files that are the pieces that make everything work. Instead of Claude opening every file in a folder trying to figure out what's relevant, it reads one small file first, then goes straight to the two or three notes it needs. On the day-to-day, this means when I'm digging into something for my UX coursework, for example, Claude pulls the two or three research summaries relevant to that topic rather than my full References folder. When I'm brainstorming a new pitch, it pulls active-projects.md and by-area.md to check what I've already covered instead of scanning every draft I've ever started. Projects and memory sound like they should do this already I mentioned this setup to a friend a while ago and they said something along the lines of "why not just use Projects or let Claude's Memory handle it?" That's a fair question, but neither replaces my vault method… Projects are a persistent workspace on claude.ai. You upload files to a knowledge base, write project instructions once, and every new chat inside inherits both automatically. Memory is separate. As of March 2026 it's available on every plan including free, and it runs a synthesis pass over your standalone chat history roughly every 24 hours to build a profile that gets injected into future conversations. Memory from Projects stays walled off from your global memory too, which is either useful or annoying depending on what you need to get done. I utilize both of them because they both do their jobs well. But Projects lives on claude.ai and I'm working in Claude Code against local files. Project instructions also load on every message rather than once, so a fat instruction block costs tokens every turn. And memory is a synthesis I don't control - it's fine for general information about me but it's not going to know where my most recent pitches currently are in the folder. Where the rate-limit savings actually come from And the routing matters more than the size of the cache Claude Code meters against a rolling 5-hour session plus a weekly cap and that pool is shared with claude.ai and Cowork, so heavy use in one drains the others. And it's measured in tokens, not messages or prompt count. Anthropic doesn't publish per-message counts because a message can be anything from a few hundred tokens to tens of thousands depending on what got loaded. The two costs to actually watch are conversation history (every turn replays the whole chat plus every file Claude has read so far) and baseline context. Anthropic's own help center notes that a complex CLAUDE.md, several MCP servers, and a large project directory can put your starting cost per message in the thousands of tokens before you've even started anything. The vault setup avoids both. My SKILL.md is around 400 tokens and the CLAUDE.md schema is maybe another 400, plus the notes Claude ends up pulling are usually 1-2k each. That means a session lands somewhere in the 2-4k range. A Project doing the same job would carry instructions plus a knowledge base easily an order of magnitude bigger, replayed on every turn. I think the routing is what really helps things here. By that I mean the one small file telling Claude which other files to open, so the big stuff never has to load in the first place if they don't need to. This wasn't the initial goal when I started using Obsidian with Claude, but I'm not complaining about the results My Obsidian vault was never designed to be a cache. I initially just wanted some help organizing the mess and keeping on top of the junk flying in. The more I hooked Claude Code into it though, the more it turned into the thing my other AI tools kept trying to be with knowledge bases and memory profiles. And I noticed hitting my rate limits less often. There was a point where I'd hit it after just two hours of doing research and synthesis in a Project, now I do the same thing in my vault without needing to load in the entire vault, and never hit the ceiling. Obsidian OS Windows, macOS, Linux, iOS, iPadOS, Android Individual pricing Free normally; $4/month for Obsidian Sync