What to Build

Build an AI agent with memory, using Mastra as the agent framework and Elasticsearch as the memory store. Your agent should answer questions it couldn't answer without remembering something: your data, your history, your context.

Start from the starter repo and pick your tier:

Tier 1 - Easy Win: Elasticsearch as long-term memory. Ground a Mastra agent on your own data. Ingest any corpus - games, recipes, docs, products, songs (sample datasets included) - and get an agent that only answers from what's really in the index, with citations, via hybrid keyword + semantic search. No timestamps needed. If you're new to agents, you'll have a working demo in under an hour.

Tier 2 - Advanced: episodic memory with time decay. Add the time dimension. Typed memories retrieved with hybrid recall (ES|QL FORK → FUSE) weighted by DECAY, so recent context outranks stale context and superseded decisions lose to their replacements. The retrieval machinery is built - your job is data with a shift or reversal in it, and the tuning: decay window, fusion weights, the query, the agent's instructions.

Bonus on either tier: use Mastra's built-in memory primitives (semantic recall and working memory) backed by Elasticsearch; a reference agent ships in the Easy Win starter. It counts in the judging rubric under "Use of Mastra," as does retrieval tuning under "Use of Elasticsearch."

Any domain is fair game: personal assistants, team knowledge, ops runbooks, recommendation engines, whatever your data makes interesting. The best demos show a before and after: the same question answered without memory vs. with it.

What to Submit

• Your project on DevPost, before submissions close (8:00 pm sharp - submit early; you can keep hacking after you submit). Include:
  • A project name and a 2-3 sentence description of what your agent remembers and why it matters
  • A link to your code (GitHub repo or fork)
  • Which tier you built (Easy Win / Advanced / both) and what data you used
• A live presentation: 1-2 minutes, hard cap. The winning format: ask the same question twice, memory-blind vs. memory-aware, with the Mastra Studio trace visible so judges can see the tool calls, the recall, and (if you used them) the memory primitives at work.