The agent itself manages its memory (RAM/cache/disk) inside its reasoning loop.
- Storage
- Cloud, or local (git-backed MemFS)
- Retrieval
- The agent pages memory across OS-style tiers (core / recall / archival) via function calls inside its reasoning loop.
- Self-host
- Self-host: moderate
- License
- Apache-2.0
- Pricing
- Free (3 agents); Pro $20/mo; Team/Enterprise custom · Freemium
- GitHub stars
- 23,554
- Last release
- 2026-05-14
- Last commit
- 2026-06-26
- First catalogued
- 2026-06-28
Strengths
- Self-editing memory
- OS-style core/recall/archival
- Git-backed MemFS
Watch out
- It's a whole runtime, not a drop-in layer — bigger commitment
Best for
- Long-running autonomous agents that manage their own memory
- Memory-first coding agents (Letta Code, the current flagship)
Benchmark results
| Benchmark | Value | Backbone | Trust | Source |
|---|---|---|---|---|
| locomo | 74 accuracy | gpt-4o-mini | Self-reported | Letta (MemGPT authors — Packer, Wooders et al.) ↗ |
Sources
- https://www.letta.com/ (vendor)
- https://github.com/letta-ai/letta (vendor)
- Letta's Next Phase — Letta Code becomes the flagship coding agent (vendor)
Last verified 2026-06-28 · updated by refresh-framework-cards