Why Your Agent Cannot Find a Document It Already Received

An agent can fail to find an uploaded document even when the upload succeeded. The document may still be processing, belong to a different scope, have an outdated revision, or fail to reach the final prompt. Diagnose those boundaries separately before concluding that the model forgot the file.

Use one document with a distinctive fictional fact, such as “Project Larch uses review code Q17.” Ask for that fact in the original session and then a new session. The exact source ID and processing state matter more than whether the answer sounds confident.

Did the application persist the source?

An attachment visible in a chat window does not establish how long the backend retains it. Record the application file ID, provider document ID, owner, and revision. Confirm that the new session resolves the same permitted source rather than creating a new empty workspace.

If the upload returned an accepted response, inspect readiness separately. A background parser can fail after the initial request succeeds. Report processing, failure, and readiness explicitly so users do not repeatedly upload the same file while it is still being indexed. Supermemory's document operations provide the relevant management API; use the documented states rather than guessing from elapsed time.

Can the retrieval layer find it directly?

Query the retrieval service without the answer-generation step. Start with the distinctive identifier, then try the user's paraphrase under the same authorized scope. Save returned source IDs and versions.

If the exact identifier works but the paraphrase fails, inspect extraction, chunk boundaries, and retrieval settings. If neither works, inspect readiness and permissions first. Increasing the number of results will not fix a source that was never indexed or is correctly excluded from the user's scope.

Did the evidence reach the model?

A retrieved passage can disappear during deduplication, summarization, or context-budget trimming. Log the evidence IDs selected for the actual model request and compare them with the retrieval response. Avoid logging raw private document text when identifiers are sufficient.

If the correct passage reached the model but the answer missed it, examine the question, surrounding instructions, and conflicting evidence. This is a different failure from document loss. Keep the generation settings fixed while investigating it.

What should a regression test include?

Keep the original file, a replacement revision, a processing failure, another user's copy, and a deleted source. Verify both evidence selection and the answer across a fresh session. A deleted record must not return through a retry or connector resync.

The persistent knowledge-base guide explains how to retain source identity across revisions. To inspect the full path in your application, start a Supermemory test with one fictional document and follow its ID from acceptance to a cited answer.