MemGPT

Architecting Agent Memory: Principles, Patterns, and Best Practices — Richmond Alake, MongoDB

Jun 27, 2025 · 17:37

Richmond Alake from MongoDB presents memory management as the key pillar for building believable, capable, and reliable AI agents, arguing that agentic systems require structured memory types—persona, toolbox, conversational, workflow, episodic, and entity—to achieve statefulness and reduce reliance on prompt engineering. He introduces MemoRiz, an open-source library implementing these memory design patterns, and positions MongoDB as a flexible memory provider with its document model and hybrid retrieval capabilities (vector, text, graph). Alake details practical patterns: storing tool schemas for scalable tool use, persisting conversation history with timestamps and recall signals, leveraging workflow memory to learn from failures, and using MongoDB's upcoming integration of Voyage AI embedding models to simplify chunking and retrieval. He connects these advances to neuroscience, citing how feline visual cortex research inspired CNNs and noting recent collaborations with neuroscientists to further agent memory research.

Stateful Agents — Full Workshop with Charles Packer of Letta and MemGPT

Apr 19, 2025 · 1:19:34

Charles Packer, lead author of the MemGPT paper and co-founder of Letta, argues that statefulness (memory) is the most important problem to solve for building useful AI agents, since LLMs are inherently stateless transformers. He presents MemGPT's LMOS (Language Model Operating System) approach, which treats memory management as a context compilation problem solved by the LLM itself using tool calling to read and write structured memory blocks. The workshop demonstrates Letta's open-source stack (FastAPI, Postgres, Python) where agents persist state on a server, enabling long-running, learning interactions without context overflow—e.g., an agent can update its core memory (e.g., correcting a user's name) or search archival memory (e.g., recalling user preferences after a reset). Packer also shows multi-agent communication via async message passing between agents running as independent services, and highlights that tools are sandboxed by default with support for Composio integrations. The session includes a live notebook exercise and a low-code UI (ADE) showing context window management and memory editing, emphasizing that true stateful agents improve continuously over time rather…