[ TRACE · BEACON ]
Research that compounds — an agent with persistent memory, shipped solo.
Solo build — agent, memory layer, full stack|Live · built for Vercel Zero to Agent Hackathon 2026
[ THE PROBLEM ]
Most AI research tools reset every session: no memory, no continuity, no sense of what you already learned. Every conversation starts at zero, and the tenth research session is no smarter than the first.
Beacon is a research agent whose memory persists — findings from one session are retrievable in the next, so research builds on itself instead of repeating.
[ HOW IT WORKS ]
[ SPAN · COMPOUND LOOP ]
Intentresearch question
Planstructured framework
Recallmemory read
Researchgenerate + validate
Commitmemory write
- The loop closes: what stage 5 writes, stage 3 retrieves next session. The agent's tenth session starts where the ninth ended.
- Structured frameworks keep research repeatable — the same question type follows the same validation path every time.
Next.js and the Vercel AI SDK drive the agent loop in TypeScript. Research follows structured frameworks rather than freeform prompting, and every session's findings are written to a vector memory store that future sessions retrieve from.
Built end-to-end solo — agent design, memory layer, and UI — and shipped for the Vercel Zero to Agent Hackathon 2026.
[ TRADE-OFFS I MADE ]
01
Persistent memory over stateless chat
Statelessness is simpler — no storage, no retrieval relevance problems, no stale-memory risk. Beacon takes the harder path because the entire premise is continuity: the value of session N is that sessions 1 through N−1 are still in play.
02
Structured frameworks over a freeform agent
A freeform agent feels more magical demo-to-demo but produces unrepeatable research. Frameworks constrain the agent to validation paths that can be trusted — and make its memory entries consistent enough to retrieve usefully later.
[ WHAT BROKE ]
INCIDENT NOTE · PENDING
Every system here has broken at least once. This section gets written from the engineer's incident notes, not generated — the honest version is coming.
[ RESULTS ]
- designed, built, and shipped end-to-end
- Solo
- deployed and usable today
- Live
[ STACK ]
- Next.js
- TypeScript
- Vercel AI SDK
- RAG
- Vector memory
- Tailwind CSS