Inspiration
Family recipe boxes are fragile archives. A card may be handwritten, stained and missing a quantity; two relatives may have different versions of the same dish. The detail that makes the recipe usable often lives with a person, not on the card. I wanted to recover that detail without pretending the original said it, and without flattening the family's versions into one.
What it does
Table Memory is a source-linked family cookbook for UGREEN AI NAS. It works on a small approved selection of cards, scans, screenshots, PDFs and voice notes. It inventories the files, fingerprints the originals, extracts each recipe with the NAS document, photo and knowledge experts, and presents a review table. Every ingredient and step keeps its source. If a quantity, temperature, time or unit is unclear, the Skill marks it for review rather than guessing.
When a family member supplies a missing detail, they approve the exact wording. Table Memory stores it in a correction record bound to the source fingerprint, recipe version, ingredient or step locator, and field. A later session reopens that record and reuses the correction only when the source, branch, locator and field still match. A changed source triggers review instead of silent reuse. The cookbook, shopping list and correction record are saved only after approval and read back for verification.
How we built it
I built a Markdown Uliya Skill that delegates to the UGOS file, document, photo and knowledge experts. The Skill defines bounded inventory, exact handoffs, version matching, structured correction records, approval, save and read-back. I also made a separate local worked example with synthetic cards to test the correction lifecycle off-NAS: first session, saved correction, second session, changed-source rejection and read-back. The proof scripts stay outside the submitted Skill ZIP.
Challenges we ran into
The hard part was not OCR; it was provenance and scope. A correction must belong to one source and one recipe version, and it must not leak into another branch. A duplicate title is not enough to establish a family variant. Saving successfully is not enough either: the saved content has to be read back and compared with what was approved. Those constraints shaped the record schema and the review rules.
Accomplishments that we're proud of
I built and reproduced a complete local correction cycle: an unclear quantity became an attributed clarification, the correction record was saved, a separate second session reopened it and reused it for the matching recipe version, a changed source was rejected, and the saved cookbook and record were read back against a frozen baseline. The submitted Skill passes the official checker with zero errors and zero warnings. The local proof is deliberately labelled as non-NAS work, so it does not overstate what has been tested.
What we learned
A useful family archive needs to preserve uncertainty, not just fill gaps. The source and the family recollection should remain visibly separate. Persistence only earns trust when the system can explain why a correction applies, and when it can decline to reuse it after the evidence changes.
What's next for Table Memory
On the assigned UGREEN NAS, I will run the full cycle: source → identify gap → human answer → approve → save → read back → reopen → reuse. I will test handwriting, transcription, save permissions and a changed source with permissioned material. Round 1 is a Markdown Skill; the local example is not a claim of UGOS or NAS execution.
Log in or sign up for Devpost to join the conversation.