Mengram vs Letta
Letta (formerly MemGPT) pioneered agent-controlled memory from UC Berkeley research. Mengram takes a different approach with 3 structured memory types and procedural learning.
What Letta does well
- Novel agent-controlled memory architecture
- UC Berkeley research-backed (MemGPT paper)
- Free and self-hostable
- Great for long-running conversations
What Letta doesn't have
- No procedural memory
- Only partial episodic memory (conversation archival)
- No self-improving workflows
- No Cognitive Profile
- Agent memory management adds unpredictability
- Limited managed hosting options
Feature comparison
| Mengram | Letta | |
|---|---|---|
| Semantic memory (facts) | ✅ | ✅ |
| Episodic memory (events) | ✅ | Partial |
| Procedural memory (workflows) | ✅ | ❌ |
| Self-improving procedures | ✅ | ❌ |
| Cognitive Profile | ✅ | ❌ |
| Multi-user isolation | ✅ | ❌ |
| Knowledge graph | ✅ | ❌ |
| MCP server | ✅ | ✅ |
| Import (ChatGPT, Obsidian) | ✅ | ❌ |
| Self-hostable | ✅ | ✅ |
| Price | Free | Free (self-host) |
When to choose what
Choose Letta if...
Long-running conversational agents where the agent should organically manage its own context and memory.
Choose Mengram if...
Structured memory with 3 types that the developer controls. Procedures evolve from failures. Cloud API + MCP + framework integrations.
Try Mengram free
No credit card. Get your API key in 30 seconds.