Midas · vs Zep
Midas vs Zep
Hosted temporal graph vs. local-first temporal tiers.
Zep stores conversations in a hosted temporal knowledge graph (Graphiti) with LLM-driven extraction. Midas keeps temporal tiers local, with deterministic ingest and span-level provenance, optimized for long-horizon recall over months of agent history.
Feature by feature
Feature
Midas
Zep
Hosting
Local-first (single SQLite)
Hosted / self-host complex
Ingest
LLM-free, deterministic
LLM extraction graph
Recall@k (LoCoMo)
0.85 (alpha)
~0.74 reported
Provenance
Source span returned
Node/edge references
Cost per memory
≈ 0 (embed only)
LLM tokens per turn
MCP
Built-in server
Not native
Operational footprint
One file
Service + DB
When to pick Zep
- You want a managed knowledge graph and accept the per-turn LLM cost.
- Your team is committed to a graph-shaped memory model.
When to pick Midas
- You want zero ops: a SQLite file you can ship with your agent.
- You want eval-first numbers and per-recall provenance.
- You don't want LLM cost on every conversational turn.
Try Midas in 30 seconds.
pip install "midas-memory[all] @ git+https://github.com/vornicx/Midas" View on GitHub Other comparisons