Great Questions AI

Inspiration

Most AI research tools retrieve documents. They can find what someone said, but they do not preserve the evolution of thought: what you believed at a point in time, what evidence changed it, and which assumptions remain unresolved.

Great Questions AI starts from a different premise: memory is not about the past. It is an engine for better future thinking.

What it does

Great Questions AI turns podcast conversations into a sourced, searchable research memory. In the current demo, a Mastra agent answers natural-language questions over a live Elasticsearch corpus and returns the episode title, date, timestamp range, and source URL for the evidence it used.

The product is organized around three questions:

1. What did I think? Recover earlier ideas in their original context.
2. What changed my mind? Compare evidence and positions across time.
3. What's next? Use unresolved assumptions to generate better research and podcast questions.

How we built it

• Mastra provides the interactive research agent, local Studio experience, REST API, and MCP server.
• Elasticsearch Serverless is the long-term source of truth. Versioned indices and aliases hold memories, relations, and answer receipts.
• Podcast transcripts are parsed into deterministic, overlapping timestamped chunks and bulk-indexed through an alias.
• Each chunk is indexed into a semantic text field. Search combines lexical relevance and semantic retrieval with reciprocal rank fusion.
• The agent is required to search before answering corpus questions and to cite title, date, timestamp, and URL.
• OpenRouter supplies the language model used by the Mastra agent.

The working demo contains six DAMA LA podcast episodes represented by 159 Elasticsearch transcript memories. We also recovered all 19 Agentic Mesh Podcast transcripts locally as the next, more chronological corpus; they are kept out of Git and are being prepared for the same deterministic ingestion pipeline.

Challenges we ran into

The hardest part was making memory auditable rather than decorative. Transcript captions are not perfectly diarized, so the agent must not invent speaker attribution. Semantic indexing also has a model warm-up cost, and a safe public release has to exclude raw transcripts and credentials while keeping the full pipeline reproducible.

We addressed those constraints with deterministic document IDs, strict mappings, explicit aliases, long first-write timeouts, exact source locators, and grounding instructions that state uncertainty when a speaker is unclear.

Accomplishments that we're proud of

• Live hybrid search over 159 timestamped podcast memories.
• Grounded answers with click-through episode citations.
• One search capability exposed consistently as a Mastra tool and through MCP.
• A graph-ready memory model that preserves historical nodes instead of overwriting them.
• A privacy-safe repository that excludes credentials and raw podcast data.

What we learned

Retrieval becomes memory when provenance and time are first-class. A strong embedding model is not enough: the system must preserve the original evidence, distinguish episode participants from verified speakers, and make every claimed change traceable to governing memory IDs.

What's next for Great Questions AI

Next we will ingest the 19 Agentic Mesh episodes, extract validated claims, ideas, questions, and predictions beside the raw transcript memories, and add explicit SUPPORTS, CONTRADICTS, REFINES, and SUPERSEDES relations. An answer-receipt index will record which memories governed each conclusion. That closes the loop: capture, extract, reflect, generate better questions, capture again.

Built With

• elasticsearch
• mastra
• node.js
• openrouter
• semantic
• typescript
• zod