comparison

MemOS vs Mem0 vs Zep vs Letta

Four ways to give an AI agent a memory. This page is written by the MemOS team, so salt accordingly — but the numbers are reproducible from the repo, and we note where competitors don't publish.

MemOSMem0ZepLetta
Where your data livesSQLite file on your machineHosted cloud (platform) or self-managed (OSS library)Hosted cloud, or community edition self-hostedSelf-hosted server (Postgres-backed)
PriceFree — MIT, no paid tierFree tier capped; $19–$249/mo platform plansFree tier; paid plans for scaleOpen source; managed cloud available
Memory modelGraph-native, temporal validity, trust scoringFact extraction + graph memory (Pro plan)Temporal knowledge graphMemory blocks + archival storage (MemGPT)
API keys / accountNoneAccount + API key for the platformAccount + API key for the cloudNone for self-hosted
EmbeddingsLocal (Gemma-300M) by defaultProvider-hostedProvider-hostedConfigurable
MCP integrationBuilt-in adapter (stdio + HTTP/SSE)Via integrationsVia integrationsAgent-server centric
BEAM-1M recall @1095.9%64.1%not published on BEAM-1Mnot published on BEAM-1M

Pricing and feature details as published on each project's website, August 2026. Mem0 and Zep also publish open-source components; the comparison above covers the hosted platforms most teams actually run.

When to pick what

Pick MemOS when…

You want memory that never leaves the machine, no account or API key, long-horizon recall at production scale, and the freedom to modify the storage layer itself.

Pick a hosted service when…

You need managed infrastructure across a team, don't want to run anything yourself, and the per-month cost is worth the zero-ops trade-off.