Mnemos

Project Story (Inspiration/What You Learned/How You Built It/Challenges):

Mnemos is an AI agent built for professionals drowning in information silos. Rather than another search tool or note-taking app, it remembers across your inbox, calendar, and documents—then takes actions under your approval.

Inspiration: The gap between what we need to recall (commitments, context, relationships) and what our tools actually help us retrieve. Search is reactive. Notes are fragmented. We built an agent that understands why you're asking.

How Built:

Hybrid retrieval at scale: Every query runs MongoDB Vector Search + Atlas Search (BM25) in parallel, merged via Reciprocal Rank Fusion, then reranked with Gemini. This beats single-modality search consistently.
Multi-agent verification: A critic sub-agent audits every generated response against cited sources—flagging unsupported claims and hallucinations before they reach you.
Graph-augmented traversal: When identifying key entities, the agent walks your memory graph (people, projects, relations) to pull context no keyword search finds.
Cost/latency transparency: Every run streams token counts, USD cost estimates, and latency—critical for production LLM apps.

Challenges:

Balancing latency (parallel retrieval phases) vs. cost (reranking passes).
Designing an adversarial critic that catches hallucinations without over-flagging.
Wiring real Gmail OAuth and send flows securely.
Streaming reasoning + citations + telemetry in real-time over SSE.