AgentStack: The Memory Layer (MeetUp)

The talk will cover the core concepts and modern architectural components of AI Agent Memory, including long-term memory, working memory, derived knowledge, memory retrieval, and memory orchestration. It will also explore practical application patterns and discuss key challenges related to consistency, scalability, and memory management in agentic AI systems.

Organizational knowledge is fragmented. Some lives in formal knowledge bases, much is buried in closed tickets, and a large share remains tacit, surviving only as experience that was never written down. So how does Alex, our digital worker, consolidate all of this into a single, self-updating knowledge layer it can reliably reason over?

The answer lies in actively ingesting from varied enterprise sources and continuously curating the data to prevent context rot and contradiction. In this session, we will show how this agentic pipeline comes together to keep every answer reliably grounded in its source, allowing Alex to truly become a trusted extension of your team.

As agentic workflows demand longer execution paths, an agent’s "memory" becomes its defining constraint. In order to understand this better we must consider the underlying architecture. How we manage and maintain this long-term cognitive state determines whether an agent successfully executes a multi-step workflow or gets trapped in an expensive, hallucinated loop.

• The Attention Tax: How the core math of transformers creates 'danger zones' in episodic memory.
• Memory Tiering & Compression: Strategies for structuring short-term vs. long-term memory to optimize for accuracy and token cost.
• Anthropics Dreaming & Memory Consolidation: An exploration of the concept of model "dreaming”, using background LLM cycles to compress and optimize agent logs into usable memory.

The goal of this talk is to move past naive vector-search dumps and toward designing sophisticated, architecture-aware memory systems based on how LLMs actually process and retain information.