SubVault – MCP server for persistent AI memory across tools
SubVault is an open-source MCP server that gives Claude, Cursor, and Copilot a shared vault. Say vault to save decisions, remember to load them back — one memory layer for every MCP client.
TL;DR
TL;DR: SubVault is an open-source MCP server that gives AI tools persistent memory. Say
vaultto save session context (decisions, people, action items). Sayrememberto load it back — in the same tool or a different one. One command install, MIT licensed.
Source and Accuracy Notes
⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.
- Project page: subvault.ai — verified
- Source repository: github.com/subvaultlabs/subvault — README read, MIT license confirmed
- License: MIT (verified via
LICENSEfile in repo) - HN launch thread: not found (2-point Show HN, no separate launch thread as of this write)
- Source last checked: 2026-09-05
What Is SubVault?
SubVault is an MCP server that acts as a persistent memory layer for AI coding tools. When you say vault in a conversation, your AI extracts and saves the session’s key facts, decisions, people, and action items to a local SQLite database. When you later say remember, SubVault retrieves the relevant entries and injects them back into context — in the same tool or a different one entirely.
The core pitch from the README:
Say vault to save a session. Say remember to load it back later, in the same tool or a different one. Your AI never starts from zero.
It works with any MCP-compatible client: Claude (Desktop and web), Cursor, VS Code Copilot, and others.
Setup Workflow
Step 1: Get an API key
Sign up at subvault.ai/signup to receive your API key. The free tier covers personal use.
Step 2: Install
curl -fsSL https://subvault.ai/install.sh | bash -s YOUR_API_KEY
The install script configures Claude Desktop, Cursor, and VS Code in one pass. It is safe to re-run.
Step 3: Use the vault and remember commands
Once installed, just speak to your AI normally:
You: vault
AI: Done. I've saved the key decisions from this session to your vault.
You: remember
AI: From your vault — Sarah owns frontend, Mike handles DevOps. Open items: update billing webhook before Friday, blocked on Stripe keys.
Alternative: One-click Claude Desktop extension
Download subvault.mcpb and double-click it to install without the CLI.
Alternative: Web Claude
Add https://mcp.subvault.ai/mcp as a custom connector in Claude.ai’s Connectors UI, then sign in.
Full setup docs at subvault.ai/docs/setup.html.
How It Works
When you say vault, SubVault extracts structured entries from your session rather than dumping a raw transcript. Each entry is tagged by kind:
| Kind | Example | | --- | --- | | Decisions + reasoning | Chose PostgreSQL over MongoDB. Billing needs relational integrity. | | People + relationships | Sarah owns frontend, Mike handles DevOps, report to Lisa. | | Action items + status | Update billing webhook before Friday. Blocked on Stripe keys. | | Project context | API at v2, migration from v1 in progress, 3 endpoints left. |
On retrieval (remember), SubVault classifies your query (“Who is Sarah?” → people, “What’s the status?” → action items), ranks results by source, recency, and relevance, and returns about 2,000 tokens — enough to be useful without filling your context window.
Data Privacy
SubVault stores entries in your own SQLite database, not on SubVault’s servers. Only explicitly vaulted items are sent; the rest of your conversation is never transmitted. Each vault is a separate database, and deletion removes it entirely.
Comparison to Alternatives
| | SubVault | Built-in memory / transcripts | |---|---|---| | What is saved | Decisions, people, action items | Full transcript | | Relevance ranking | Yes, by source and recency | No | | Context window impact | Fixed ~2,000 tokens | Grows with every session | | Cross-tool memory | Yes | No | | Data ownership | Local SQLite | Vendor-locked |
FAQ
Q: Does it work offline? A: Yes. The vault lives in a local SQLite database on your machine. No cloud dependency for storage.
Q: Which AI tools are supported? A: Any MCP-compatible client — confirmed working with Claude Desktop, Claude.ai web, Cursor, and VS Code Copilot.
Q: Is there a self-hosted option?
A: Not yet documented in the README. The server endpoint (https://mcp.subvault.ai/mcp) suggests a hosted service, but the vault itself is a local SQLite file you control.
Q: How does deduplication work? A: Items are hashed on save. Vaulting the same thing twice stores it once.
Q: Is the source code open? A: Yes, the repo is at github.com/subvaultlabs/subvault and the license is MIT.
Practical Evaluation Checklist
- [ ] Install via
curl -fsSL https://subvault.ai/install.sh | bash -s YOUR_KEY - [ ] Configure Claude Desktop and verify MCP connection
- [ ] Run a coding session, say
vault, then start a new session and sayremember - [ ] Verify saved entries appear (people, decisions, open threads)
- [ ] Test cross-tool retrieval: save in Cursor, recall in Claude Desktop
- [ ] Check SQLite file is created locally (
~/.subvault/or similar) - [ ] Verify deletion removes the database
Conclusion
SubVault solves a real problem with long-term AI coding: every new session starts from scratch. By extracting structured memory into a shared vault, it lets you pick up exactly where you left off — across tools, without context-window bloat. The install is one command, the MIT license keeps it open, and the local SQLite storage means your data stays yours.
If you regularly switch between Claude, Cursor, and Copilot on the same project, SubVault is worth a look.
Related Posts
ai-setup
Recall – Persistent Memory for Claude Code via MCP Hooks
Recall gives Claude Code a permanent memory store that survives session restarts and context compaction. Four hooks capture and restore context automatically — with cloud SaaS or self-hosted options.
2/28/2026
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026