linkloom is a local-first, default-read-only personal knowledge assistant for Markdown and Obsidian vaults. It is designed as a real open-source portfolio project, not as a collection of Agent demos or a fixed folder-template generator.
It serves people who want materials, understanding, actions, and reflection to form a long-lived loop without manually maintaining a complex knowledge system. It must work for an existing vault first, while also supporting people who are gradually building a new one through ongoing input and review.
The planned user journey is simple:
Can Read -> Can Find -> Can Connect -> Can Organize -> Can Act
-> Modify Only After Confirmation
linkloom will help users:
- search and answer questions from notes with sources;
- discover explainable relationships between notes;
- identify duplicate, isolated, weakly tagged, or disorganized notes;
- recover tasks, unresolved questions, and next actions;
- preview and confirm any rename, move, rewrite, or metadata change before it touches a real vault.
Milestone 1's Read-Only Vault Scanner is implemented and independently reviewed against a synthetic sample vault. It safely indexes existing note structure without modifying source notes. Search, relation suggestions, health diagnostics, action continuity, and all real-vault writes remain future work.
The next planned product slice is a read-only Vault Profile: a human-readable portrait generated from the Scanner index. It is not implemented yet.
The confirmed product and planned Python package name is linkloom. The local
repository folder is still named vault-steward; renaming that directory is
not part of this planning change.
- Local-first by default.
- Read-only by default.
- Synthetic sample vaults before real user vaults.
- Evidence paths for answers, relations, diagnoses, and extracted actions.
- Read tools and write tools remain separate.
- Dry-run, exact confirmation, audit evidence, and rollback before real writes.
- No silent delete, merge, rewrite, rename, or move.
linkloom does not require one directory tree. Instead, it should support this user lifecycle across many possible vault structures:
materials enter
-> understanding and knowledge sedimentation
-> connections between themes, questions, and notes
-> goals, tasks, or experiments
-> real-world feedback and reflection
-> reviewed knowledge returns to the vault
Later "self-growing" behavior means improving this visible and controllable loop from evidence and user-confirmed rules. It never means free-form autonomous rewriting of a user's knowledge base.
- docs/PRODUCT_ROADMAP.md: canonical six-milestone user roadmap.
- SPEC.md: product contract and non-goals.
- DEV_SPEC.md: implementation gates mapped to the roadmap.
- AGENTS.md: hard rules for AI contributors.
- docs/README.md: documentation navigation map.
- docs/MULTI_AGENT_ROLES.md: Planner, Worker, Reviewer separation.
- docs/SPEC_DRIVEN_DEVELOPMENT.md: required development workflow.
- docs/PROJECTS_REVERSE_ENGINEERING.md: development patterns learned from reference projects.
- docs/MARKET_LANDSCAPE.md: adjacent tools and differentiation.
- docs/GIT_WORKFLOW.md: restore-point and commit-boundary rules.
- docs/ACCEPTANCE_CHECKLIST.md: task-level acceptance checks.
No feature implementation may start until its SPEC, acceptance criteria, and independent review gates exist and the human has approved the active planning boundary.