Submit projectSubmit

THE SOTA COMPARISON

Mem0 vs Zep vs Letta for AI agent memory

Compare extracted facts, temporal context graphs and stateful agents, with verified open-source options, hosted plans, SDKs and releases.

By SOTA · Updated 2026-11-07 · Researched with AI assistance; facts checked against the linked sources · Not a hands-on benchmark · How we work

The short answer

Mem0 adds fact extraction and retrieval to an application. Zep supplies temporal context graphs while you keep your agent framework; Graphiti is its open-source graph framework, and Zep Community Edition is unsupported. Letta manages persistent agents with Git-backed MemFS memory. Choose by how much of the memory and agent runtime you want to adopt, then test with your own conversations.

Decision points from the linked documentation
DecisionMem0ZepLetta
What it isMemory layer: add conversation messages, then search for relevant factsContext engineering platform: business data, documents and conversations in temporal Context GraphsStateful agent platform: persistent identity, memory, tools and conversation history
Open source and licencemem0ai/mem0: Apache-2.0; hosted Platform adds capabilities beyond OSSGraphiti: Apache-2.0; Zep uses a proprietary graph engine. Community Edition is deprecated and unsupportedCurrent runtime in letta-ai/letta-code: Apache-2.0, excluding brand assets; older V1 API server is retired
Self-hostingRun the library in your app or a Docker server; configure your own componentsSelf-host Graphiti with your own graph database and surrounding application; Zep BYOC is an Enterprise optionLocal CLI runtime or self-hosted App Server; self-hosted agents need manual backups and cannot use chat.letta.com
Hosted free planFree: first 30 days include 10,000 adds and 1,000 retrievals/month with model calls covered; bring your own LLM key for 50,000 adds and 10,000 retrievals/monthFree: 10,000 credits/monthFree: 3 stateful agents; bring your own model keys or external coding plans
Paid self-serve entry (USD/month)Usage-based overages with a monthly cap you set; Custom pricing for higher limits and advanced featuresFlex: $125/month; 50,000 credits, 5 projects, 5 Context MCP seats and 600 requests/minutePro: $20/month including $20 in model and remote-sandbox credits; up to 20 agents. Teams Pro: $20/seat/month
Memory modelExtracted facts with metadata and embeddings; your app chooses what enters the promptEntities and relationship facts with validity timestamps for historical and point-in-time retrievalMemFS: Git-backed Markdown files; root memories stay in context, indexed directories are read when needed
SDK languagesPython and JavaScript/TypeScriptPython, TypeScript/JavaScript and Go; Graphiti is a Python frameworkCurrent Agent SDK: JavaScript/TypeScript, with cloud, local and remote backends
Integration examplesOfficial guides for LangChain, CrewAI, LlamaIndex and Vercel AI SDKLangGraph, CrewAI, Google ADK and Vercel AI SDK integrations; keeps your framework's runtimeHosted MCP server for other agents; local messaging channels include Slack, Telegram and Discord
Latest checked package or releasePyPI mem0ai 2.2.1: September 25, 2026PyPI zep-cloud SDK 4.1.0: November 2, 2026; graphiti-core 0.31.2: October 20, 2026GitHub letta-code v0.34.44: November 7, 2026; current agent runtime
Limits to budget or checkHosted use beyond the allowances is usage-based; your own LLM key adds provider costs. OSS also needs model, embedding and storage infrastructureFlex automatically tops up at 20% balance: $25 per 10,000 credits; Free has lower-priority processingAdditional hosted credits can be purchased; bring-your-own model usage is billed by that provider. MemFS has no vector index by default

How each approaches memory

Mem0 sits between your application and model. Send messages to add, search before a model request, and decide which results to include in the prompt. By default it stores extracted facts rather than a verbatim transcript. Zep combines conversations, documents and business data in temporal Context Graphs. Facts carry validity bounds that applications can use for historical retrieval; those timestamps do not prove the underlying claim is true.

Letta persists the agent's identity, tools, configuration and conversations as well as memory. Its current documentation describes MemFS: a Git-backed memory repository projected into Markdown files. Root memories enter the system prompt each turn, while directories with their own MEMORY.md index are read when needed. MemFS has no semantic or vector index by default. This is the current documented model, so older memory-block and archival-storage descriptions should not be assumed to describe the current runtime.

Open source and hosted products

Mem0's repository is Apache-2.0 licensed and can run as a library or Docker server. The hosted Platform shares core extraction and retrieval logic but adds capabilities beyond OSS. Self-hosting means choosing and paying for your model, embedder and storage components.

Zep announced the end of maintenance and releases for Community Edition on April 2, 2025; the repository now marks it deprecated and unsupported. Its open-source effort focuses on Apache-2.0 Graphiti. Graphiti requires your own graph database and surrounding user and conversation management, while Zep uses a proprietary graph engine. Zep offers Enterprise deployment inside your VPC; Graphiti is not a self-hosted copy of the hosted service.

Letta's current source is in letta-ai/letta-code. Its Apache-2.0 licence excludes the repository's brand assets. The letta-ai/letta repository retains the retired V1 API server on its archive branch. Current self-hosting uses the local runtime or App Server, with manual backups and no access through chat.letta.com. Letta's September 23, 2024 announcement assigned the Letta name to the agent framework and kept MemGPT for the original agent design pattern.

Pricing and limits

These are published USD monthly prices checked November 7, 2026. Mem0 now lists Free and Custom, replacing the older Hobby and Starter comparison. It counts adds and retrievals separately and meters usage beyond the allowance against a cap you set. After its initial model-covered period, plan for your own LLM key and its provider bill. Zep charges one ingestion credit per 350 bytes or part; retrieval and storage are unmetered. Flex automatically tops up at 20% balance for $25 per 10,000 credits.

Letta now prices individual Pro at $20/month with $20 in credits for models and remote sandboxes, plus a 20-agent limit. Teams Pro is $20 per seat per month, with shared agents and permissions; Enterprise is custom. Free supports three agents with your own model keys or an external coding plan. Additional hosted credits are available to purchase. Estimate ingestion, retrieval, model and tool costs for your workload.

How to choose by situation

Consider Mem0 when you already have an agent runtime and want to add fact storage and retrieval through an add/search loop. Consider Zep when your application needs temporal relationships across conversations and business data while retaining its existing framework. Consider Graphiti when you want to build and operate the graph system yourself. Consider Letta when you want persistent agents with their own identity, tools, conversations and editable memory files, and can adopt its runtime and SDK. These are architectural choices based on the documentation; validate them with representative conversations from your application.

What we did not test

This comparison is based on official pages, repository licence files and package or release pages read on November 7, 2026. We did not install the software, run agents, benchmark retrieval, measure latency or evaluate memory accuracy. The release row identifies specific packages and repositories; it does not assign a software version to the hosted services. No performance, popularity or overall ranking is claimed.

Sources & method

Official documentation was checked on 2026-11-07. Selection advice is our assessment. Interfaces and requirements may change; check the linked source for the version you intend to use.

How SOTA reviews projects · Suggest a correction · Download the evaluation checklist

Explore the projects