Memory-first design¶
Waymark is a memory-first companion, not a generic document-search tool. That distinction drives every design decision. The core loop is always:
capture messy thought -> structure memory card -> store locally -> retrieve with sources -> reflect over time
Non-drift rules¶
These rules keep the product from drifting into "yet another RAG app":
- Memory before automation. Manual capture is the foundation; AI is an enhancement.
- Capture before import. Documents enrich the journey — they don't replace your own captured memories.
- Local SQLite backbone before vector search. A simple, inspectable relational store comes first.
- App-only mode before local AI setup. Everything works with no model installed.
- Explicit confirmation before model downloads, file scans, OCR, or background indexing. Nothing heavy happens by surprise.
- Every generated answer must cite saved memories or imported sources.
- Decisions are first-class objects, not just tagged notes.
What this means in practice¶
- Private by default. All data is local. Local AI is opt-in and runs through Ollama on your machine; no remote calls.
- Models are never downloaded for you. Setup prints the exact
ollama pull ...commands and lets you run them. - Folders are never scanned automatically. You always name the file or folder, and imports are preview-first.
- AI output is always a draft. Generated titles, summaries, tags, and reflections are shown for confirmation before anything is saved.
- Answers are grounded. If no saved memory or source matches, Waymark says so and invites you to capture or import more context — it does not fabricate.
Guided first, commands second¶
The guided Textual interface is the primary experience: calm, warm, and keyboard-driven, framed around attention and documentation coverage rather than scoring your life. Every guided action has a direct CLI equivalent for scripting and automation.