Midas · vs Letta (MemGPT)
Midas vs Letta
Agent runtime with memory vs. memory primitive for any agent.
Letta (formerly MemGPT) is an agent runtime: memory is bundled with the agent loop. Midas is the opposite — a memory primitive (SDK + MCP server) that any agent runtime can call. You bring the loop; Midas brings durable, traceable recall.
Feature by feature
Feature
Midas
Letta
Scope
Memory primitive (SDK)
Full agent runtime
Long-horizon recall@k
0.95 / 0.85 (LME-s / LoCoMo)
Not eval-published
Integration
pip install, MCP server
Adopt the runtime
Ingest
LLM-free
Self-edits via LLM
Persistence
SQLite, portable
Postgres / managed
Temporal tiers
Working / episodic / archival
Working / archival
Selective forgetting
First-class API
Manual
When to pick Letta
- You want a batteries-included agent runtime and accept its opinions.
- You're prototyping with Letta-native tools.
When to pick Midas
- You already have an agent loop (LangGraph, custom, MCP client).
- You want memory as a library, not a framework.
- You need durable recall measured on public benchmarks.
Try Midas in 30 seconds.
pip install "midas-memory[all] @ git+https://github.com/vornicx/Midas" View on GitHub Other comparisons