MemGPT/Letta - Long-term Memory Management
Basic Information
- Product Name: Letta (formerly MemGPT)
- Development Company: Letta Inc.
- Country/Region: USA
- Official Website: https://www.letta.com/
- GitHub: https://github.com/letta-ai/letta
- Research Paper: https://research.memgpt.ai/
- Type: Stateful AI Agent Platform / Long-term Memory Management
- License: Apache-2.0
- Funding: Seed round investment from renowned VCs like Felicis
Product Description
Letta (formerly MemGPT) is a platform for building stateful AI agents, treating memory as a first-class citizen of the agent's state. Its core innovation is the "LLM as OS" paradigm—where LLMs actively manage their own memory, context, and reasoning loops, much like an operating system manages memory. Agents achieve infinite memory capacity within fixed context window limits through a hierarchical memory structure.
Core Features/Characteristics
- Hierarchical Memory Structure:
- Core Memory: Persistent labeled context blocks (goals, preferences, roles, etc.), always injected into agent prompts
- Conversational Memory: Management and summarization of conversation history
- Archival Memory: Out-of-context memories stored in a database, retrieved on-demand via search
- External Files: Accessible external data sources
- Agent Self-Managed Memory: Agents actively decide what to keep in immediate context
- Conversations API: Supports agents maintaining shared memory across multiple users
- Stateful Memory Runtime: Transparent and developer-controllable memory management
- Editable Memory Blocks: Both developers and agents can edit memory content
- Letta Evals: Open-source evaluation framework (released October 2025)
- Model-Agnostic: Supports various LLMs
Business Model
- Open Source Edition: Apache-2.0, core platform free
- Letta Cloud: Hosted service (paid)
- Enterprise Edition: Enterprise-level features and support
Target Users
- Developers needing to build stateful AI agents
- Personal assistant and conversational AI application developers
- AI applications requiring cross-session persistence
- Scholars researching long-term memory and agent architectures
Competitive Advantages
- Innovative "LLM as OS" architectural design
- Memory as a first-class citizen, not an add-on
- Agents actively manage their own memory (not passive storage)
- Strong academic background (MemGPT paper widely cited)
- Letta Code ranked #1 on Terminal-Bench
Comparison with Competitors
| Dimension | Letta/MemGPT | Mem0 | Zep |
|---|---|---|---|
| Core Concept | LLM as OS | Memory Layer API | Temporal Knowledge Graph |
| Memory Management | Agent Self-Managed | Automatic Extraction | Graph Construction |
| Architecture | Hierarchical Memory | Hybrid Storage | Graphiti Graph |
| Best Use Case | Stateful Agents | Personalized Applications | Enterprise Complex Reasoning |
Relationship with OpenClaw Ecosystem
The "LLM as OS" concept of Letta/MemGPT provides important reference for the memory system design of OpenClaw. OpenClaw's personal AI agents require a similar hierarchical memory architecture—core memory storing key user preferences and identity information, archival memory storing complete history. Letta's agent self-managed memory approach is particularly suitable for OpenClaw's autonomous agent scenarios.