Inspiration

We built Gauntlet because modern work and personal life generate too much scattered information across chats, docs, meetings, and calendars. We wanted an AI memory system that acts like a second brain: it remembers context continuously, connects related facts, and helps users act before things slip through the cracks.

What it does

Gauntlet ingests continuous micro-updates (“micro-dumps”) and turns them into searchable memory with semantic embeddings and a knowledge graph.
It can:

Recall context from past conversations and events
Link people, projects, topics, and tasks
Trigger proactive alerts for time-based commitments
Draft actions (like replies or scheduling) with a confirm-before-send safety gate

How we built it

We used a hybrid architecture:

Neon Postgres for durable cloud memory
IndexedDB for offline/local resilience
Gemini text-embedding-004 for semantic retrieval
A knowledge graph with nodes (person, project, event, topic) and relational edges
Better Auth + Google OAuth for secure per-user access and Google Workspace integrations

Server functions are scoped by context.userId to keep user data isolated and secure.

Challenges we ran into

Designing reliable continuous ingestion instead of one-shot memory dumps
Keeping entity resolution accurate (e.g., same person mentioned multiple times)
Balancing proactive automation with trust and control
Managing integration token handling securely for external tools

Accomplishments that we're proud of

Built a unified memory stream that works across personal and professional contexts
Implemented semantic recall plus graph relationships for richer context understanding
Added proactive alert generation based on detected date/time entities
Established a clean integration path for external tools (Google first, Slack/Jira next)