Fix: Claude Project Files Not Updating (How to Force Refresh Knowledge)

I ran into this one myself last month. I updated a pricing sheet inside a Claude Project I use for client proposals, kept chatting in the same conversation, and Claude kept quoting the old numbers back to me — numbers I'd deleted from the file twenty minutes earlier. My first thought was that I'd uploaded the wrong file. I hadn't. This is an actual, well-documented behavior in how Claude Projects handles knowledge, not a one-off glitch, and once I understood why it happens, fixing it took about thirty seconds.

If you're stuck watching Claude confidently give you outdated answers, or you're seeing two versions of the same file sitting in your project knowledge, here's what's actually going on and the exact steps that get Claude reading the current version again.

Why Is Claude Ignoring My Uploaded Files?

The short answer: Claude isn't ignoring your files on purpose — it's most likely still working off a cached read from earlier in the conversation, or it's pulling the wrong chunk out of a knowledge base that now has more than one version of the same document sitting in it.

Here's the part that trips most people up. Project knowledge in Claude isn't a live folder that Claude re-scans every time you send a message. When you upload a file, Claude processes it once at upload time. On paid plans, once your project's total knowledge grows past what fits directly in the context window, Claude automatically switches to Retrieval Augmented Generation (RAG) mode — instead of reading every file in full each time, it uses a search tool to pull out only the pieces it thinks are relevant to your current question. Anthropic explains this shift in their own RAG for Projects documentation.

That's genuinely useful for keeping large projects fast, but it also means two things that aren't obvious until you hit them:

1. A conversation that started before you updated a file may keep referencing whatever got pulled into its context earlier — updating the file mid-chat doesn't retroactively refresh what's already been loaded.
2. If you upload a revised file without removing the old one first, Claude's knowledge base now technically contains both, and its retrieval step can grab content from either one.

Neither of these is you doing something wrong. It's just how the current system is built.

Does Claude Projects Automatically Update Files?

No — and this is the core limitation worth understanding before you go looking for a "sync" button that doesn't exist. Claude Projects has no automatic re-indexing on a schedule, and there's no live connection between a file on your computer and the copy sitting in a project's knowledge base. Every update is manual. If you edit a document locally, Claude has no idea that happened until you go back into the project and re-upload it yourself.

This also applies to connected sources like Google Drive or GitHub repos — I've seen multiple reports of connected repos showing as "Connected" in the project while actually serving stale or no content at all, which forced people back to manually uploading files one at a time as a workaround. If you're relying on a connector rather than direct uploads, don't assume it's silently staying current — check it the same way you'd check a manually uploaded file.

How Do I Force Claude to Read New Project Files?

This is the actual fix, in the order I'd try them:

1. Start a brand-new chat inside the project

This is the single biggest fix and the one most people skip. Don't keep typing in the same conversation where Claude already gave you the outdated answer — that thread has context baked in from before your update. Open a fresh chat inside the project. New chats pull from the current state of the knowledge base rather than whatever got cached earlier in an existing one.

2. Delete the old file before uploading the new version

Don't just add "pricing-sheet-v2.md" next to "pricing-sheet.md" and hope Claude figures out which one matters. Go into the project's knowledge base panel, remove the outdated file completely, confirm it's gone, and only then upload the replacement. Having two versions of the same document in there is exactly what causes Claude to give inconsistent answers or seemingly "create duplicate files" — it's not creating anything, it's just retrieving from whichever version its search step happens to land on.

3. Confirm the upload actually landed

Open the knowledge base panel after uploading and check that the file is listed with today's date and the size you'd expect. Uploads occasionally fail silently on a slow connection or a file that's too large, and it's easy to assume it worked when it didn't.

4. Try converting the file format

If a file has been in the project a while and keeps producing odd or partial answers even after a clean re-upload, try converting it to plain .txt instead of .md, .docx, or .pdf. In my own testing, plain text files with simple line breaks get parsed and chunked more predictably than heavily formatted Markdown or Word files, especially once RAG mode is active. It's not a guaranteed fix, but it's a quick thing to try if formatting-heavy files are the ones giving you trouble.

5. Use a direct prompt to force a fresh retrieval

Inside your new chat, don't just ask your question normally — explicitly tell Claude to go check the knowledge base again. Something like:

"Before answering, search the project knowledge base for the current version of [filename] and confirm the date or details you're pulling from."

This nudges Claude to actively re-run its retrieval step rather than lean on anything it might already have in context, and it also gives you a way to catch the problem immediately — if Claude quotes back old details after you've asked it to confirm, you know the file itself still needs fixing rather than the conversation.

6. As a last resort, start a new project

If a project's knowledge base has accumulated a lot of overlapping, renamed, or duplicate files over time, it can genuinely get messier than it's worth cleaning up. Starting a fresh project with a clean, current set of files takes a few minutes and sidesteps whatever confusion built up in the old one.

Why This Isn't Really a "Bug" So Much as a Workflow Gap

Claude currently has no way to write back to project files either — it can read what you give it, but if you ask it to update a document in your knowledge base directly, it can't; you have to download what it generates and re-upload it yourself. Combine that with no auto-sync on the read side, and you get a system where staying current is entirely on you. It's worth building "delete old file, upload new file, start new chat" into a habit if you update project knowledge regularly, the same way I'd recommend building any repeatable check into your routine — I cover a similar habit-building approach in my guide to Claude Cowork's Record a Skill feature.

FAQs

Why is Claude ignoring my uploaded files? Usually it isn't ignoring them — it's either still using context from before you updated the file within the same conversation, or your knowledge base has an old and a new version of the same document, and Claude's retrieval step pulled from the wrong one. Starting a new chat and removing outdated files before re-uploading fixes most cases.

How do I force Claude to read new project files? Delete the outdated file from the project's knowledge base, upload the current version, confirm it shows up correctly in the panel, and start a new chat rather than continuing the one where you saw the outdated answer. You can also directly ask Claude to search the knowledge base and confirm what it's referencing before it answers.

Does Claude Projects automatically update files? No. There's no scheduled re-indexing and no live sync between a file on your device and what's stored in the project. Every update — including through connectors like Google Drive or GitHub — needs to be manually verified, since connected sources can also show as connected while actually serving outdated content.

Why does Claude seem to create duplicate files in my project? Claude isn't generating duplicates on its own — this usually happens when a revised file gets uploaded without removing the original, leaving two versions of essentially the same document in the knowledge base for Claude to retrieve from.

---

If you're managing a project's knowledge base for ongoing client or business work, it's worth pairing this fix with a quick monthly cleanup of anything outdated or duplicated. For a broader look at how Claude compares to other assistants on context and memory handling, see my ChatGPT vs Gemini vs Claude comparison, and for the general context-loss issue on the ChatGPT side, I covered a similar fix in why ChatGPT memory stops working.