AI Agents SF #12 — AI Memory
AI Memory talks:
(1) "Own Your Memory" by Gauri Nagavkar, Lead Evangelist at MemVerge, creators of MemMachine.
(2) "Building Token-Efficient Memory Systems for Production: Benchmarks, Tradeoffs, and Design Decisions" by Livia Ellen, Growth Engineer at mem0
(3) "Semiont: A human+AI collaborative context engineering and agent memory platform" by Adam Pingel of AI Alliance
(4) Lightning Talk
"Cure Agent Amnesia: Graph-Powered Memory for AI Agents" by Guy Lubovitch, Head of Customer Engineering @ FalkorDB
"Building a Knowledge Graph with Memory & Vectors" by Vasilije Markovic of Cognee
Talk details and speaker bios:
"Own Your Memory" AI agents are powerful, but fundamentally amnesiac — each interaction starts from scratch, with no reliable way to retain or reuse what was learned before. Many current approaches rely on context windows or custom pipelines that are fragmented and difficult to control. In this talk, we’ll explore what it really means for AI systems to have memory, and why memory, not just context, is emerging as a core layer for building useful, long-running agents. We’ll also look at how to design systems that explicitly store and retrieve memory over time, with MemMachine as one example of such a memory layer.
Gauri Nagavkar is a Lead Evangelist and Developer Advocate at MemVerge, creators of MemMachine, focused on building AI systems with persistent memory that learn and evolve over time.
"Building Token-Efficient Memory Systems for Production: Benchmarks, Tradeoffs, and Design Decisions" Most memory benchmarks measure accuracy in isolation, but in production, accuracy must balance token cost, latency, and scale. This talk breaks down what LoCoMo, LongMemEval, and BEAM actually measure. We’ll explore the tradeoffs between accuracy and cost and how we built Mem0’s open-source memory algorithm to achieve strong results across benchmarks while using 3-4x fewer tokens per query.
We will cover the key design decisions behind it: single-pass extraction, entity linking, and multi-signal retrieval.Livia Ellen is a Growth Engineer at Mem0 (YC S24), the memory layer for AI agents. She also writes In Context, a weekly Mem0 column on context engineering and memory systems.
"Semiont: A human+AI collaborative context engineering and agent memory platform"
Adam Pingel, Head of Knowledge Engineering, AI Alliance
"Cure Agent Amnesia: Graph-Powered Memory for AI Agents" Agents forget everything once a conversation ends. We'll show how FalkorDB + memory solves this with graph-structured memory, replacing flat-file storage with nodes and edges that preserve real relationships.
Guy Lubovitch, Head of Customer Engineering @ FalkorDB
"Building a Knowledge Graph with Memory & Vectors"
Vasilije Markovic, Founder of Cognee
Sponsored by: Neo4j
We are looking for more talks on the fundamental questions of Agentic AI platforms, framework and protocols, especially whether a platform layer will hold. Comment with proposals or email us!