Stateful agents with self-managed, layered memory that learn and improve over time
Key Features
Multi-Agent Single Agent Human-in-the-Loop Streaming Async Support Short-Term Memory Long-Term Memory Shared Memory Plugin System Custom Tools MCP Protocol A2A Protocol Structured Output API Server Self-Hosted Cloud Hosted
Community Feedback
Strengths
- Best-in-class memory management
- Pioneering LLM-as-OS paradigm
- Research-backed (UC Berkeley)
- Strong documentation
- #1 model-agnostic on Terminal-Bench
Weaknesses
- Memory management complexity
- Architecture transition (MemGPT to Letta V1)
- Learning curve
Letta (MemGPT) Details
| Organization | Letta Inc (UC Berkeley spin-off) |
| Organization Type | Company |
| Funding | Seed |
| Category | Framework |
| Subcategory | Single agent |
| Deployment | Hybrid |
| Primary Language | Python |
| Runtime | Python 3.10+ |
| License | Apache-2.0 |
| Commercial Use | Permissive |
| Install Command | pip install letta |
| GitHub Stars | 21,736 |
| GitHub Forks | 2,295 |
| Release Cadence | Biweekly |
| Maturity | Stable |
| Pricing Model | Open core |
| Free Tier | Apache-2.0 licensed core is free |
| Self-Hosted Free | Yes |
| Cost Model | free + LLM costs |
| Community Size | Large (22k stars) |
| Community Activity | Very active |
| Sentiment | Positive |
| GPU Required | No |
| Research Date | 2026-03-24 |
| LLM Providers | OpenAI, Anthropic, Google, local models |
| API Keys Required | LLM provider API key |
Use Cases
- Long-running conversational agents
- Agents that learn and improve over time
- Persistent personal assistants
- Knowledge management with agent memory
When to Use
Best for: Long-running agents needing persistent, self-managed memory
Avoid when: Simple stateless agent workflows or rapid prototyping
Original data from HuggingFace, OpenCompass and various public git repos.
Check out Ag3ntum — our secure, self-hosted AI agent for server management.
Release v20260328a