AI Memory Keeper

Your AI conversations, distilled into a private, cited memory on your NAS: decisions, fixes and facts from any assistant, answered with sources, never leaving your hardware.

Comment

Inspiration

I solve real problems in long conversations with several AI assistants. The useful parts ("we chose X because Y", "the fix was Z") stay locked in each vendor's chat history, get lost when I switch tools, and cannot be searched across assistants. I have been building a local-first memory for AI conversations for a while; this hackathon is a chance to make the idea native to a NAS, where my files already live.

What it does

AI Memory Keeper is a UGOS Skill with three workflows:

  • Capture: drop exported conversations (ChatGPT, Claude, Gemini and others) into /AI-Memory/inbox and say "Process my new chat exports." Each conversation becomes one dated note with decisions, solutions, facts and open questions, linked to the original export. Processed files are tagged so they are never processed twice; moving them to an archive needs confirmation.
  • Recall: ask "What did we decide about the backup schedule?" The answer comes only from saved notes and cites the note and the export. If nothing was saved, the Skill says so instead of guessing.
  • Context card: "Prepare my context for a new assistant" builds a one-page card (standing decisions, preferences, known fixes) to paste into any assistant.

How I built it

The Skill is Markdown only, following the Uliya Skill standard: SKILL.md describes three roles (Collector, Distiller, Librarian), each working on the actual output of the previous step, using only NAS capabilities: Files find/tag/organize, Documents read/extract/answer, Knowledge save text. References define the note format, the test cases and the disclosure.

Challenges

Keeping the memory trustworthy: nothing is invented, every statement points to its source, contradictions are shown with dates instead of being merged, and secrets in exports are replaced by "secret omitted". No file is ever deleted and nothing leaves the NAS.

What's next

Round 2: map every step to the exact UGOS CLI commands and test extraction quality on real ChatGPT and Claude exports on a UGREEN AI NAS.

Built With

Share this project:

Updates

Submission history