Building a Subconscious for AI Agents: Passive Awareness Across Five Communication Channels

One of the core problems in multi-agent AI systems is context fragmentation. Each agent only knows what it was explicitly told or what it can read in shared files. But a lot of relevant context lives in communication channels—emails, messages, meetings—that no agent is watching continuously. We built a “subconscious” layer to solve this.

What the Subconscious Does

Every 30 minutes, a background indexer pulls recent messages from five channels and extracts structured concepts—people, decisions, deadlines, action items, concerns—and stores them in a searchable SQLite database. Any agent can then query this database to get ambient awareness of what’s been happening without having to read every raw message.

The five channels are: a personal email inbox, a business email inbox, a work collaboration platform, a team chat, and a mobile messaging app. Together they represent the full surface area of daily communication.

The Extraction Pipeline

Each channel has its own collector that handles authentication and data format. The collectors run in parallel; failures are isolated so a broken channel doesn’t kill the whole indexing run. All messages from the last 24 hours get concatenated and passed to a single LLM call that extracts concepts in a structured JSON format.

Each concept has a body, a source type, a timestamp, and optional metadata. The concept bodies are indexed with FTS5 full-text search so agents can query with natural language terms like “client meeting” or “deadline” or a person’s name.

Why SQLite Instead of a Vector Database

We evaluated vector databases for this use case and decided against them. The queries agents run are mostly keyword-based—looking for mentions of a specific person, project, or date. Vector similarity search adds latency and complexity without meaningfully improving recall for this pattern. SQLite FTS5 is fast, requires no running server, works offline, and the database file is just a file you can inspect or back up like any other.

How Agents Use It

When any agent starts a task that might benefit from recent context—drafting a message, preparing a report, triaging a request—it can query the subconscious with relevant keywords before proceeding. The results give it passive awareness: things it didn’t explicitly learn but that are now available because other communication happened. This is the same thing a well-briefed human assistant does automatically.

The Privacy Consideration

All processing happens locally. No messages are sent to external services for indexing. The concept extraction uses a local or session-scoped LLM call, and the resulting database lives on-device. This is a non-negotiable property of the system—communication data doesn’t leave the machine.