-
-
Ai-memory-vault logo
Inspiration:
ChatGPT has memory for ChatGPT. Claude has memory for Claude. Cursor has its own context. But users do not really own one memory that works everywhere. AI Memory Vault was inspired by that problem: people repeat the same project details, preferences, goals, and decisions across many AI tools. We wanted to build a local-first memory layer that users control.
What it does:
AI Memory Vault stores personal AI memories locally and makes them reusable across tools. Users can: Save memories from the dashboard or browser extension Capture ChatGPT, Claude, Gemini, Copilot, or custom AI chat context Review and approve memories before they are used Search approved memories semantically Insert Vault context into new AI chats Export/import memory data Connect AI tools through an MCP server The core idea is simple: One memory. Many AIs. User owns it.
How we built it
We built AI Memory Vault as a local-first monorepo. The backend uses Python FastAPI with SQLite for memory storage. Search uses a lightweight local vector-style retrieval mode, with support for sentence-transformers/Qdrant-style expansion. The frontend is built with React, Vite, and Tailwind CSS. It includes a dashboard for memory review, search, import/export, suggestions, timeline, and extension setup. The browser extension is built for Chrome and works on ChatGPT, Claude, Gemini, Copilot, and custom AI chat pages. It can preview chat text, generate memory suggestions, save full transcripts, search the Vault, and copy selected context back into an AI prompt. We also added an MCP server so tools like Claude Desktop, Cursor, Windsurf, and other MCP-compatible clients can retrieve and manage memories.
Challenges we ran into
The hardest part was making memory useful without becoming automatic surveillance. We designed the extension so it does not scrape in the background. The user must explicitly click to save or use context. Another challenge was browser extraction. ChatGPT and other AI apps do not expose full conversation history through a stable API, so the extension can only capture what is loaded in the page. We improved selectors and extraction logic to capture loaded user/assistant messages more reliably. We also hit local development issues with SQLite inside OneDrive folders, which caused disk I/O errors. We fixed this by moving active local data storage to a safer app-data location.
Accomplishments that we're proud of
We are proud that the MVP is not just a CRUD memory database anymore. It now has the start of a real product experience: Save memory from any AI Review before use Use approved context anywhere Local-first by default No external LLM required Browser extension as the main workflow MCP support for AI tool integration The project proves that personal AI memory can live outside individual AI platforms.
What we learned
We learned that the real value is not “another memory database.” The value is portability. Users do not want to manually export and import everything. They want memory to follow them naturally across ChatGPT, Claude, Cursor, custom tools, and future AI products. We also learned that trust is essential. Memory should be visible, editable, exportable, and approved by the user before retrieval.
What's next for Ai-memory-vault
Next, we want to improve automatic memory suggestions and make the browser extension smoother. Planned improvements: Better full-chat capture across AI sites Smarter memory extraction from long conversations One-click approval workflow Better timeline and provenance views Stronger local encryption Desktop app packaging More MCP client setup guides Support for Gmail, Slack, Teams, and VS Code Optional local LLM summarization
Log in or sign up for Devpost to join the conversation.