Where vault-cortex is heading over roughly the next year — and what it deliberately won't do. Priorities shift with real-world usage, and this document is updated when they do.
Last updated: 2026-09-02.
vault-cortex exists so that anyone can give their AI tools ready access to their personal knowledge base, task management, and durable memory on top of an Obsidian vault they own. A vault holds years of accumulated knowledge — knowledge that should be reachable from wherever the person works, in a shape AI tools can navigate, and findable by meaning as well as by exact keywords. The planned work clusters around those goals: making setup accessible, deepening the task and memory layers, and hardening the project for long-term use.
The foundations the planned work builds on:
- Vault reads, writes, and surgical edits, with hybrid search (keyword + semantic + cross-encoder reranking) and the About Me/ memory layer with entry-level recall
- A vault-wide task layer — structured queries and one-call updates in both Tasks-plugin formats (emoji and Dataview)
- OAuth 2.1 alongside static token auth
- One self-contained image for every deployment path — the MCP server and Obsidian Sync in a single container, runnable by any OCI runtime (Docker, Podman, nerdctl)
- A guided CLI (
npx vault-cortex) that walks through setup and handles configuration, upgrades, restarts, and logs - One-click Render and Railway deploys, with browser-based Obsidian Sync sign-in when no token is configured
One-click deploys (Render, Railway) and browser-based Obsidian Sync sign-in already remove most of the command line from remote setup. Next:
vault-cortex initstarts the container into setup mode when no Sync token is available, so the browser flow covers the CLI path too- Platform-aware hints on the setup page (where to find your token on Render vs Railway)
- Friction fixes as early adopters work through the guides
A vault that works well with AI needs structure — memory files, session protocols, a task board. The companion project vault-onboarding scaffolds that structure through a conversational interview, with per-client instructions for Claude Code, Cursor, Perplexity, Copilot, and others. Expect the two projects to pair more tightly: onboarding output that is ready to serve through vault-cortex on day one.
- Checklist progress on task listings, so filtered board reads show how far along each card is
- Safer structured edits — atomic markdown table row operations instead of find-and-replace splices
A v1 release that MCP clients can build on:
- A final naming sweep across tool inputs and outputs
- Unknown input parameters rejected instead of silently ignored
- A written compatibility policy for what future versions may change
Write support for canvas, base, and other text files, closing the gap where remote vaults can read but not edit them.
- A software bill of materials (SBOM) published with each release
- Fuzzing for the markdown parser layer
- Graph queries over the vault's wikilink graph (multi-hop paths, neighborhoods)
- MCP resources as a third primitive alongside tools and prompts
- Large-vault performance: result pagination, incremental startup indexing
- Native Windows execution — Docker is the supported runtime on every platform
- A hosted service — vault-cortex is self-hosted by design; your vault stays on infrastructure you control
- Becoming an Obsidian plugin, or requiring one — the server works on vault files directly and stays useful even where Obsidian isn't running
- Note systems other than Obsidian — the server is deliberately Obsidian-first: its link resolution, properties, and plugin-format support match what Obsidian does