Session Overview
Discover why treating AI agent memory as an append-only event log solves the debugging and auditability limits of vector databases.
In this talk, Naga Durga Rao Dindi explores the perils of mutable agent memory in autonomous AI systems. When an autonomous agent makes a bad decision or books the wrong option, traditional vector databases can't explain what context existed at that exact moment. Learn how event sourcing with Apache Kafka allows you to record every prompt, tool call, and state mutation, enabling full context replay, time travel debugging, and sandbox forking.
What You'll Learn
- The flaw of vector lookups: Why upserts and nearest-neighbor retrieval destroy operational history
- The event-sourced agent model: Recording prompts, tool payloads, and observations as immutable events
- Anatomies of an agent event: Session keys, offsets, causation tracing, and timestamps
- Projections and working context: Balancing compacted facts, episodic summaries, and semantic indexes
- Time travel debugging: Reconstructing exact agent belief state at any historic offset
- Sandbox forking: Replaying recorded events in isolation to test new models and policies
- Tradeoffs and production costs: Storage bounds, replay latency, idempotency, and data privacy
Chapters
- 00:00 — Introduction & why agent decisions are hard to debug
- 02:22 — The problem with vector DBs: Upserts destroy history
- 05:06 — Rebuilding agent memory as an immutable Kafka log
- 10:00 — Anatomy of an agent event: Offsets, payloads & causation
- 12:13 — Context projections: Working context, episodic summaries & vector views
- 14:14 — Compacted facts vs. memory summaries
- 15:25 — Time-travel debugging & sandbox forking for policy testing
- 18:44 — End-to-end architecture: Agent runtime to stream projections
- 21:22 — Engineering tradeoffs, retention costs & key takeaways
