MemGPT
An AI memory management system, now renamed Letta, that abstracts memory into a hierarchical architecture comprising a main context and external storage. This design allows the AI to autonomously manage the loading and unloading of memories, effectively extending its usable context beyond fixed window limits.
Core Facts
MemGPT is a memory management solution that breaks through LLM context window limitations by simulating the virtual memory mechanism of an operating system.
Mainstream approaches to long-term memory management for current AI characters include writing key interaction events as structured summaries into an external memory store (similar to the design of MemGPT) and dynamically retrieving and injecting them into context during subsequent conversations.
Projects such as LangChain's Memory module and MemGPT have conducted extensive exploration in the direction of long-term memory.
Timeline (last 90 days)
Letta (MemGPT) has 25,091 GitHub stars, offers a web app and CLI, supports OSS server self-hosting, starts at $0 then Pro at $20/month, and rewrites memory in the background via sleep-time agents.
Memory layering ideas similar to Devin's are also reflected in systems such as MemGPT and Letta.
Research from MemGPT, Reflexion, and others found that having agents explicitly generate experience summaries after task completion and write them back to memory can significantly improve cross-task transfer capability.
The experience reuse mechanism is conceptually similar to a combination of RAG and few-shot prompting; it deliberately avoids model fine-tuning and shares the same lineage as external memory solutions such as MemGPT (now Letta).
On the recall-focused portion of the StuLife benchmark, JAZ's invoke outperforms Letta (MemGPT) by 8% at half the cost.
Long-term memory evaluation benchmarks (such as LongMemEval and the MemGPT benchmark) assess five dimensions: single-hop information extraction, cross-session reasoning, knowledge updating, temporal reasoning, and abstention judgment.
MemGPT treats the limited context window as memory and external storage as a disk, managing information flow through an explicit paging mechanism.
Projects such as LangChain's Memory module and MemGPT have conducted extensive exploration in the direction of long-term memory.
MemGPT was proposed by UC Berkeley in 2023, building a hierarchical memory system for LLMs similar to an operating system's virtual memory.
Mem0 focuses on providing a persistent personalized memory layer for AI applications, while LangChain Memory offers a variety of memory management strategies ranging from buffer memory to entity memory.
2 more timeline events
All Facts (17)
MemGPT is a memory management solution that breaks through LLM context window limitations by simulating the virtual memory mechanism of an operating system.
80%VerifiedMainstream approaches to long-term memory management for current AI characters include writing key interaction events as structured summaries into an external memory store (similar to the design of MemGPT) and dynamically retrieving and injecting them into context during subsequent conversations.
75%VerifiedProjects such as LangChain's Memory module and MemGPT have conducted extensive exploration in the direction of long-term memory.
65%VerifiedMem0 focuses on providing a persistent personalized memory layer for AI applications, while LangChain Memory offers a variety of memory management strategies ranging from buffer memory to entity memory.
65%VerifiedMemGPT has been renamed to Letta; it simulates the virtual memory mechanism of an operating system, allowing agents to manage information that exceeds the context window capacity within a limited context window.
65%Unverified斯坦福大学的生成式智能体（Generative Agents）研究和Letta（原MemGPT）等项目也在探索跨会话持久记忆和系统级学习方向
85%UnverifiedUser profile construction involves the storage and retrieval of long-term memory; typical frameworks such as MemGPT and Mem0 adopt a mechanism of persisting historical conversation summaries or structured tags and retrieving them for injection at the start of new conversations.
60%UnverifiedMemGPT has been renamed Letta and proposed the concept of a hierarchical memory architecture, dividing memory into three levels: working memory, short-term memory, and long-term memory.
60%UnverifiedLetta (MemGPT) has 25,091 GitHub stars, offers a web app and CLI, supports OSS server self-hosting, starts at $0 then Pro at $20/month, and rewrites memory in the background via sleep-time agents.
50%UnverifiedMemory layering ideas similar to Devin's are also reflected in systems such as MemGPT and Letta.
50%UnverifiedResearch from MemGPT, Reflexion, and others found that having agents explicitly generate experience summaries after task completion and write them back to memory can significantly improve cross-task transfer capability.
50%UnverifiedThe experience reuse mechanism is conceptually similar to a combination of RAG and few-shot prompting; it deliberately avoids model fine-tuning and shares the same lineage as external memory solutions such as MemGPT (now Letta).
50%UnverifiedOn the recall-focused portion of the StuLife benchmark, JAZ's invoke outperforms Letta (MemGPT) by 8% at half the cost.
50%UnverifiedLong-term memory evaluation benchmarks (such as LongMemEval and the MemGPT benchmark) assess five dimensions: single-hop information extraction, cross-session reasoning, knowledge updating, temporal reasoning, and abstention judgment.
50%UnverifiedMemGPT treats the limited context window as memory and external storage as a disk, managing information flow through an explicit paging mechanism.
50%UnverifiedMemGPT现已更名为Letta
50%UnverifiedMemGPT was proposed by UC Berkeley in 2023, building a hierarchical memory system for LLMs similar to an operating system's virtual memory.
50%