The website artbitrarily limits the context window to 50 prompts (25 user prompts + 25 responses) which is ludicrusly low in most cases and completely negates models with large context windows. It leads to awkward fixes like downloading the conversation, manually removing all the blocks (which is extremely time consuming) then feeding it back to a new conversation as a single prompt - and repeat, until one hits the actual…
Environment where the instability occurs: Public Mind File encoding: UTF-8 File sizes tested: 500c - 10,000c File content: Plain text with markdown patterns and tabulations Issue: Text files uploaded as text/plain spontaneously change to application/octet-stream, becoming unreadable by the Mind Assistant. Evidence: Assets like t0HL57jS2c4NJf and s4gSUmjIKEHhpo flip between readable and unreadable states without user intervention. Files are always UTF-8 encoded .txt files. Hypotheses: Background re-scanning changes detection post-upload AI…