Skip to main content

ZeroDB vs Letta

Letta is a platform for building stateful agents with their own memory system. Here's what's actually different, based on Letta's own public documentation, fetched on 2026-11-02.

DimensionZeroDBLettaChecked
Memory modelFour explicit tiers — working, episodic, semantic, procedural — each memory tagged with memory_type, with promote() to move a memory up a tier.Two named tiers: "Core Memory Blocks" (persistent in-context knowledge) and "Archival Memory" (external, out-of-context store).docs.letta.com/memory, fetched 2026-11-02 — documentation uses "Core Memory" and "Archival Memory" terminology, not working/episodic/semantic/procedural.
Knowledge graphrelate()/graph() + hybrid use_graph recall, confirmed live this session (401 on real routes, 404 control).No knowledge-graph feature found in the fetched documentation.docs.letta.com/overview and docs.letta.com/memory, fetched 2026-11-02 — no graph/relationship feature mentioned.
Deployment optionsHosted, zero-auth instant provisioning via a single POST (npx zerodb-cli init), confirmed by a real run this session. ZeroDB Local for offline Docker Compose use.Letta Agent SDK, a web app + CLI, Letta Cloud, and a self-hosted Letta App Server — the agent harness itself is open source.docs.letta.com/overview, fetched 2026-11-02.
PricingLetta's overview page did not surface specific pricing tiers or dollar amounts in what was fetched this session — left as an open question rather than guessed.
Agent-reasoning quality / tool-use benchmarksLetta is positioned around full agent state (memory + reasoning loop), not just a memory store — a fair comparison there needs an agent-level benchmark, which was not run in this session.

What this page does and doesn't claim

Letta solves a broader problem than memory alone — full stateful-agent infrastructure, including the reasoning loop itself. This page compares specifically the memory layer, which is the part of Letta's own public docs we fetched and checked.

ZeroDB — part of AINative Studio