Stenobird

Episode

From Data Models to Mind Models: Designing AI Memory at Scale

Podcast
Data Engineering Podcast
Published
Feb 22, 2026
Duration seconds
3467
Processing state
processed
Canonical source
https://www.dataengineeringpodcast.com/agentic-memory-design-and-application-episode-502
Audio
https://op3.dev/e/dts.podtrac.com/redirect.mp3/serve.podhome.fm/episode/f6ff0caa-931b-4c08-bfdd-08dc7f5cd336/639073981323893202146bcc19-cf93-4fa2-9e96-70bb5d041afc.mp3
JSON
/v1/public/podcasts/data-engineering-podcast/episodes/from-data-models-to-mind-models-designing-ai-memory-at-scale
Markdown
/podcast/data-engineering-podcast/from-data-models-to-mind-models-designing-ai-memory-at-scale.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/data-engineering-podcast/episodes/from-data-models-to-mind-models-designing-ai-memory-at-scale/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/data-engineering-podcast/from-data-models-to-mind-models-designing-ai-memory-at-scale.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Designing AI memory requires moving beyond simple vector stores to a multi-layered architecture of graph and vector layers. This discussion explores how to implement agentic memory that distinguishes between session and long-term state while maintaining multi-tenant isolation.

Topics

  • Agentic Memory
  • Vector Databases
  • Graph Databases
  • Multi-tenancy
  • Cognitive Science
  • AI Infrastructure
  • Knowledge Engineering
  • LLM Orchestration

Highlights

  • Main idea: Agentic memory must differentiate between short-term session state and long-term persistent knowledge
  • Practical takeaway: Use a hybrid graph and vector approach to enable complex relationship retrieval and temporal searching
  • Failure mode: Avoid naive summarization or uncontrolled fine-tuning as substitutes for structured memory systems
  • Security insight: Implement physical isolation and multi-tenancy to prevent agents from accessing unauthorized data silos
  • Future trend: The emergence of 'pseudo-languages' or structured SQL-like commands for more efficient multi-agent communication

Chapters

  1. 1:00 Introduction to Knowledge Engineering: Vasilije Markovic discusses his transition from data engineering to cognitive science and the inspiration for building memory engines.
  2. 5:30 Defining Agentic State: The necessity of providing agents with a state representation to allow for continuity and data exchange across sessions.
  3. 10:00 Long-term vs. Session Memory: Exploring use cases for persistent memory, such as maintaining user profiles and breaking down data silos.
  4. 18:30 Optimizing Embedding Retrieval: Technical considerations for optimizing embeddings and managing large-scale retrieval for agentic workloads.
  5. 23:10 Temporal Relevance and Decay: How to model the temporal aspect of memory and handle information that becomes outdated over time.
  6. 27:20 Advanced Search Architectures: Implementing multi-modal search across graph and vector stores to find temporal and relational data.
  7. 31:30 Multi-tenancy and Security: Architecting isolated memory stores to ensure agents cannot cross-contaminate or access sensitive company data.