I am building this as a personal Mac expense tracker that feels like chatting with a sharp finance assistant, not filling a spreadsheet by hand. The app lets me describe expenses in normal language, confirms the structured transaction before saving, stores the real data locally in SwiftData, and keeps a CSV export available for Numbers or Excel.
The bigger idea is that the assistant should accumulate useful context over time.
Alongside transactions and analytics, it now has an app-managed Markdown memory layer inspired by OpenClaw-style local agent memory and Andrej Karpathy's llm-wiki idea.
The in-app LLM can maintain memory.md, daily notes, and a small wiki inside a repo folder I choose in Settings, while the Swift app validates every file write path.
Core guardrails:
- No transaction is saved without a confirm card.
- SwiftData is the source of truth.
- CSV is write-only export.
- OpenRouter keys live in Keychain.
- The LLM may maintain Markdown memory, but only inside the configured memory folder and only through validated
.mdoperations.