One Simple System Gave All My AI Tools a Memory. Here's How.

OpenBrain’s memory becomes genuinely useful when it gains a “human door”: a visual interface that reads and writes the same underlying database table that AI agents use through MCP. The core idea is architectural—keep the database table as the single source of truth, then expose it through two native pathways. Agents query and update rows via MCP for reasoning and automation, while people interact through fast-scanning web or mobile views designed for browsing, editing, and conflict-checking. That dual-access design matters because it eliminates the usual failure modes of agent workflows: stale data, laggy sync layers, and the friction of endless chat scrolls when the real need is to glance, sort, and update.

The system starts with a structured table inside the user’s OpenBrain setup (often built on Supabase). Instead of treating the database as a text-only “keyhole,” the approach layers a lightweight visual over the same table. The agent side stays unchanged: it can read, write, and reason across the table contents through MCP, and any updates appear immediately in the human interface. The human side is intentionally built for scanning rather than conversation—think search bars, category filters, calendar views, and highlighted exceptions—so the interface supports quick decisions and direct edits without forcing users to re-enter information through chat.

This design also clarifies why many existing chatbot or MCP-based setups feel limiting. Chat interfaces are great for session-based Q&A, but they don’t provide the persistent, glanceable surfaces needed for daily work. With a visual layer, an agent can still do the heavy lifting—like cross-referencing schedules to surface conflicts—but the user can see the result in a calendar or dashboard and update details directly on a phone.

To build these “extensions,” the workflow is: define the table schema, connect the agent via MCP, then generate a small web app that renders the table in a mobile-friendly format. The transcript highlights using an AI (Claude or ChatGPT) to generate the app code from a natural-language description of the desired view, then hosting it on Vercel to get a live URL that behaves like an app without app-store friction. The key promise is that the visual layer can be deployed quickly and cheaply because it’s not a separate data product—just a view into the user’s own table.

From there, the use cases shift from “remembering notes” to “proactive advantage.” Household knowledge becomes a structured capture pipeline: paint colors, appliance details, warranty dates, and service history get logged during ordinary conversations, then surfaced through search and categories. Professional relationship management becomes time-bridged: the agent can detect neglected contacts by scanning logged interactions and context, then the interface can show which relationships need attention this week. Job hunting becomes cross-stream reasoning: the agent links postings, contacts, conference notes, and prior conversations to generate warm-introduction leads and flag follow-up windows before momentum dies. Across all examples, the system’s value comes from linking events across months and categories humans rarely cross-reference.

Long-term, the approach is positioned as future-proof because it depends on agent-readable data rather than UI scraping or platform-controlled access. As models improve, the same logged patterns and tables should yield better conflict detection, anticipation, and automation—while the user retains control through the visual “human door.”