Player not loading? Watch on YouTube
This comparison examines how Mem0, the managed Zep service, Cognee and Letta retain information beyond an AI agent's current context window. The speaker separates stored memory from conversational context and compares retrieval methods, infrastructure changes and the cost of writing new memories.
Mem0 extracts facts into a vector store and checks whether to add, update or delete them. The presenter recommends it for adding recall to an existing agent, while noting limits in its flat memory model for relationship and history queries. The managed Zep service uses a temporal knowledge graph with timestamps for when facts were true and when the system learned them. The video describes Graphiti as its open source engine and says Zep's self-hosted edition is deprecated.
Cognee builds a knowledge graph and adjusts memory weights after feedback. The speaker describes a single-Postgres deployment option, but warns that incorrect corrections can spread through memory. Letta gives agents editable core memory and external archival storage. Its paging tools add latency, depend on the model's function calling ability and require adopting Letta's runtime.
The closing advice is to test with your own conversations. The presenter questions vendor benchmark comparisons and stresses memory extraction and update costs alongside retrieval speed. His recommendation is to start with simple recall, then add graph reasoning or agent-controlled memory when observed failures justify the extra complexity.
