Claude Project Knowledge Not Working? Check What Is Loaded

If Claude project knowledge is not working, confirm that the chat belongs to the right project and test whether Claude can locate one exact passage in the current source file. Then separate a missing file, a retrieval miss, and an instruction conflict. Repeating “read everything” makes those failures harder to distinguish.

While you investigate, fenbs can hold the approved requirements and verified troubleshooting results outside the failing conversation. That gives the next assistant a starting point without treating the last wrong answer as project truth.

What is the fastest useful test?

1. Open the intended project and confirm the affected chat is inside it. Write down the project name and the visible file list.
2. Choose one current document and one short passage whose answer you can verify yourself.
3. Ask Claude to find that named document and quote the relevant sentence before answering a narrow question.
4. Compare the quoted passage with the actual file. Record whether the file was missing, the passage was wrong, or the interpretation was wrong.
5. Change only the likely cause and repeat the same question in a fresh project chat.

For example, use a fictional studio document called “Maple approval notes, final” with a clearly marked delivery method. Ask for the delivery method and its supporting sentence. Do not use a broad request such as “draft the whole proposal” as your first test; too many things can go wrong at once.

Claude’s project management help↗ (opens in a new tab) distinguishes project knowledge from instructions and the project’s name and description. Put essential context in the actual knowledge or instructions, rather than relying on the description to communicate it to Claude.

Is the right source actually available?

Inspect what is present in this project, not what you remember uploading somewhere. Look for similar project names, duplicate file names, and attachments that were supplied only to an earlier conversation. Open the source yourself so you know which wording the assistant should find.

Check the content as well as the filename. A document named “final” can still contain an old draft. A scanned page may look readable to you while providing poor text for a narrow search. Try a small text document containing the approved passage to distinguish a source-format problem from a broader project problem.

Keep the original intact during the test. If the short text file works, you have narrowed the issue; you have not proved that the whole original document is usable. Record which page or section still needs attention before returning to a larger task.

For background on the product, Claude’s projects overview↗ (opens in a new tab) describes knowledge and project instructions. Interfaces can differ, so use the controls visible in your project instead of assuming every screenshot applies.

Did Claude search for the wrong passage?

Claude documents project retrieval, a search process that selects relevant uploaded material rather than loading every document at once. Its project retrieval guidance↗ (opens in a new tab) explains the behavior and when it applies. A file being stored and its relevant paragraph being found are different checkpoints.

Narrow the request with the document’s exact name, a section heading, and the question you need answered. Ask Claude to identify the source before applying it. “Find the delivery instructions in Maple approval notes, final, then state the approved method” is easier to check than “Use the knowledge base.”

If Claude produces a plausible answer without a supporting passage, mark the result unverified. Ask it to say when it cannot find the evidence. Do not keep restating the expected answer and then count its agreement as proof that source retrieval has recovered.

Once you have a repeatable successful question, keep the document name and verified finding in fenbs. Future assistants can start from that finding and still check the original source when needed.

Are old uploads or conflicting instructions changing the answer?

Compare the available documents for competing answers. An early estimate and an approved specification may both mention the same delivery method. Make their status clear: which document is current, what replaced the old choice, and whether an unresolved comment still needs a person’s decision.

Then inspect project instructions and the current chat request. A request to “reuse last month’s proposal” may pull the answer away from this project’s approved terms. Resolve the wording after checking the source; do not simply add a longer instruction that repeats the same conflict.

If the symptom is a capacity warning rather than a wrong answer, use the separate project knowledge capacity guide. If the affected instructions are repository files in Claude Code, use CLAUDE.md troubleshooting; that is a different place to investigate.

When is this a service or support problem?

If uploads fail, the knowledge panel will not open, or unrelated chats also fail, check Claude’s service status↗ (opens in a new tab). Match the affected product and time rather than assuming an old incident explains today’s result. A status page is a useful check, not a diagnosis for every incorrect answer.

When the issue persists, follow Claude’s support instructions↗ (opens in a new tab). Include the exact action, error text, timestamp with time zone, device or browser, and a harmless reproduction document where possible. Remove private client material from anything you share unnecessarily.

Keep separate notes for connection problems. If Claude cannot reach an outside tool at all, the Claude Desktop MCP troubleshooting guide covers that connection. An uploaded document search miss does not by itself show that an outside connection is broken.

How to do this with fenbs

1. Create a free account, free for one person with no credit card, and add the project whose source material you are checking.
2. Add a task for the failure. Include the narrow question, document version, observed response, and the next test in its plan.
3. Record the current requirements and the decisions a person has approved. Add a context note with the verified source finding and remaining uncertainty.
4. Connect a supported assistant through Integrations. MCP is a standard way for an AI app to use an outside tool; approve the connection through browser sign-in.
5. Ask the assistant to read project context before working and save progress after the check, including what passed and what remains unresolved.

fenbs does not repair Claude’s document retrieval or automatically import its project files. It is useful for current work, agreed decisions, and handoff context; keep long reference documents in the document tool that suits them. The AI memory topic covers other ways to keep context available between sessions.