Building a Personal LLM Wiki: Ensuring AI Memory Reliability with Immutable Sources and MOC Structure

The existing note app had over 900 notes but lacked inter-document connections, causing AI to mislink context. To resolve this, a personal wiki was built based on Andrey Karpathy's LLM Wiki pattern.

The core concept is the separation of raw sources and curated wiki pages. Raw sources are kept immutable (read-only) to preserve the record at the time, while the curated pages written by AI contain current conclusions. This role distinction prevents AI from confusing speculation with established facts. Additionally, a MOC (Map of Content) hierarchy was introduced to structure content into Home-Domain-Individual Page levels, and importance descriptions were added to each link to prevent the creation of orphan pages.

Operational rules prohibit moving or renaming raw sources to prevent broken links. Processing status is managed in a separate ledger to maintain source immutability, and a code-first verification procedure is used to address the staleness of work policies. The wiki records the 'why', while code verifies 'if it is still the case'.

After two months of operation, approximately 100 pages (including 12 analysis pages) were generated, and the migration of the existing 900 notes was completed. In a case involving a point cancellation API error, the AI used wiki search to accurately distinguish different root causes for the same error message, such as order number length mismatch (17 digits vs. 14 digits) or N+1 queries. However, memory aging and user input omissions remain limitations, and designing an automatic collection loop to address these is the next step.