MLOps.community Building an Agenetic AI Memory Framework
28 snips
Oct 14, 2025 Biswaroop Bhattacharjee, a Senior ML Engineer at Prem AI, dives into the fascinating world of AI memory systems. He discusses Cortex, an innovative framework inspired by human cognition, highlighting how it manages long-term and multimodal memories. The conversation challenges the boundaries of agentic memory, weighing the necessity of forgetting and the implications of memory consolidation. Biswaroop also shares insights into hierarchical collections, retrieval techniques, and the pursuit of integrating vision and audio for a richer AI memory experience.
AI Snips
Chapters
Books
Transcript
Episode notes
Relationship-First Memory
- AMEM forms relationships between memory chunks by summarizing and tagging each insertion before linking to similar past memories.
- This creates a knowledge-graph-like memory where LLMs name and score relationships during insert and retrieval.
Human Memory As Background Process
- Biswaroop compares AI short-term memory to chat windows and places long-term memory as background dumping and consolidation.
- He likens human long-term processing to background tasks that gradually interconnect memories with metadata.
Forget-or-Archive Tradeoff
- Forgetting and spaced repetition are core human memory dynamics but it's debatable whether agents should mirror forgetting.
- Agents can either prune outdated info or archive everything and rely on selection during retrieval.
