Memory Agent

Organizations with deep operational expertise carry most of it in people's heads. When experienced specialists leave, that knowledge walks out with them. New team members spend weeks interrupting colleagues for guidance on edge cases, regulatory nuances, and process judgment calls — slowing productivity, increasing error rates, and burning out the senior people they rely on. There was no scalable way to capture what veterans knew and make it continuously available to everyone else.

Built a dual-agent AI system: one agent listens to subject matter experts and captures their knowledge, the other surfaces it to associates on demand. Experts speak naturally in a real-time voice session — the capture agent extracts structured insights, categorizes them against a domain taxonomy, and builds a knowledge graph linking related concepts. Associates call the assistant agent with questions and receive cited, expert-backed answers drawn from the living knowledge base. Every session is monitored end-to-end for quality and trace analysis.

• 01

  Defining the taxonomy before building the agent made extraction significantly more reliable — the model needs a schema to extract into, not just a prompt to extract from.

• 02

  Real-time audio changes the UX contract entirely. Users expect the agent to feel present, not transactional. Streaming partial transcripts rather than waiting for final output was critical for perceived responsiveness.

• 03

  Observability traces revealed extraction failures that testing alone never caught — specifically on multi-topic utterances. Instrumenting before launch is not optional on voice agents.

• 04

  The knowledge graph visualisation became the most-used feature by managers — not by design, but because it answered the question everyone actually had: what do we know, and what are we still missing?