MemGPTFramework giving LLMs long-term memory and self-managed context beyond fixed token limits

MemGPT is an open-source framework designed to address one of the fundamental constraints of large language models: their fixed context window. Originating from research at UC Berkeley, the project introduced the idea of treating an LLM's limited context like an operating system manages limited physical memory, using paging and hierarchical memory tiers to give models the appearance of a much larger, persistent memory. The core approach borrows directly from operating system design. MemGPT distinguishes between in-context memory (the tokens currently in the model's prompt window) and external storage held outside the context. The LLM itself is given function-calling tools that let it decide when to move information between these tiers — for example, saving important facts to long-term storage, retrieving relevant past information, or editing its own core memory. This self-editing behavior is what allows agents to maintain coherent, evolving state across long conversations or documents that far exceed a single context window. The framework is aimed at developers building conversational agents that need persistent memory of users and prior interactions, as well as those working on document analysis over corpora too large to fit in context. By managing recall memory, archival storage, and a working context, MemGPT enables agents that can reference details from much earlier in an interaction without the developer manually engineering retrieval pipelines for every case. MemGPT works with both proprietary models such as those from OpenAI and locally hosted open models, and it integrates with vector databases and other storage backends to persist memory between sessions. The project later evolved and is closely associated with Letta, a company and platform that continues development of the underlying stateful-agent concepts, offering a server and tooling around the original ideas. Its main strengths are conceptual clarity and a concrete, reusable pattern for long-term memory that goes beyond naive retrieval-augmented generation. The trade-offs are typical of agent frameworks: the self-editing memory loop relies heavily on the model's function-calling reliability, which can vary with smaller or local models, and the additional memory-management steps add latency and token overhead. As an evolving open-source project, its naming, APIs, and surrounding ecosystem have shifted over time, which can make documentation and versioning a moving target.