From MemGPT to Letta
In October 2023 Charles Packer and the team at the UC Berkeley Sky Computing Lab publish the paper “MemGPT: Towards LLMs as Operating Systems”, proposing a direct analogy between the memory management of an operating system and that of a language model. On 23 September 2024 the project is renamed Letta and becomes the technological foundation of Letta Inc., a spin-off dedicated to commercialising the agentic stack. The code remains public under the Apache 2.0 licence.
The rebrand reflects a broadening of scope: while MemGPT focused on virtual memory for LLMs, Letta positions itself as a complete framework for persistent stateful agents, with management of users, tools, models and roles.
Memory hierarchy
The core idea introduced by the MemGPT paper is the memory hierarchy: an LLM context is treated as a limited memory tier — analogous to RAM — flanked by larger but slower tiers. Letta distinguishes three levels: the model’s context window (immediate active memory), the archival storage (long-term memory queryable via semantic search) and the recall memory (log of past interactions).
The agent itself, via function calling, can decide when to move information between tiers: storing a fact in archival storage, retrieving it through a query, updating the core memory block that represents information always present in the context. This mechanism allows overcoming the context window limits while maintaining coherence over long sessions.
REST API and Agent Development Environment
Letta exposes a REST API for creating and managing agents: each agent is a persistent entity identified by an ID, with state serialised in a relational database (PostgreSQL or SQLite). State includes memory, tool configuration and message history.
The Agent Development Environment (ADE) is a web interface for inspecting an agent’s internal state: core memory visualisation, archival storage navigation, tool call debugging. The ADE is designed as a developer tool during agent design and behaviour analysis.
Model-agnostic
Letta is compatible with the main providers — OpenAI, Anthropic, Google, Mistral, Ollama, LM Studio, vLLM — and with open weight models executed locally. Memory management is implemented at framework level and requires no native support from the model: any LLM capable of function calling can serve as the core of a Letta agent.
Link: letta.com