Oh
Install Oh
Theme
Appearance

Comparisons

Compare Oh

Compare where each memory system runs, what it stores, and how agents use it. Oh keeps linked records and their history in one local SQLite file.

01

At a glance

Six ways to give an agent memory, checked September 28, 2026.
ImageOhMem0SupermemoryZep and GraphitiLettaClaude memory tool
What it isLocal memory library and CLIMemory for each app userMemory API with document connectorsTemporal knowledge graph of factsAgents that manage their own memoryClaude tool for memory files
LicenseMITApache-2.0 SDK; hosted PlatformMIT repository; hosted PlatformGraphiti Apache-2.0; Zep hostedApache-2.0Part of the Claude API
How memories get inExplicit record writesA model extracts facts from messagesModel extraction; Platform connectorsModel extracts entities and factsThe agent rewrites its own memory blocksClaude writes files through tool calls
Local storage
SQLitelibSQL optional
Self-hosted SDKOr hosted Platform
Self-hosted binaryOr hosted Platform
Graphiti graph storeZep uses cloud or VPC
Self-hosted App ServerOr Letta Cloud
Your application’s storageClaude API calls still required
Matched run in Oh’s repositoryNot applicable30 questions in the MemEval harness60-question LongMemEval-S pilotNoneNoneNone
Supermemory Platform connectors include Google Drive, Gmail, and Notion. Graphiti supports Neo4j, FalkorDB, and Neptune. See the dated sources below for each product.
02

Oh vs Mem0

Mem0 is a memory layer for applications: its add call distills messages into facts scoped by user_id, so a product can remember each of its end users, either on the hosted Platform or through the self-hosted Apache-2.0 SDK. Oh keeps an agent’s own working memory instead, with source links on every fact. The page reports the one matched run so far: 30 questions in a third-party harness, too few to support a claim that either system is better.

03

Oh vs Supermemory

Supermemory is a hosted memory API: connectors sync Google Drive, Gmail, Notion, and more into containers partitioned per user, and usage is metered in credits on paid tiers. The page reports the matched 60-question LongMemEval-S pilot, where Supermemory scored higher by a margin the pilot could not separate from a tie, and what each system does that the other does not.

04

Zep, Letta, and Claude’s memory tool

Graphiti, Zep’s open-source framework, builds a temporal knowledge graph: a model extracts entities and facts from each episode, each fact keeps the window when it was true, and each traces back to the episode it came from. It runs on Python with a graph database; Zep is the managed service. Pick it to ask what was true at a given time. Oh stores what you write without a model rewriting it, in one SQLite file, and needs no graph database.

Letta, formerly MemGPT, is an agent harness and app server whose agents rewrite their own memory blocks as they work. Pick it to run stateful agents on Letta. Oh is a library you call from your own agent.

Anthropic’s memory tool lets Claude create, read, edit, and delete files under /memories, while your application runs each operation against storage it controls. It needs no extra package on the Claude API. The tool defines no schema, source links, or digests for those files, and Oh ships no adapter for it.

Oh has run no matched benchmark against these three. Oh’s benchmark records quote published LoCoMo figures for Zep and Letta and a LongMemEval figure for Zep, from setups Oh’s runs do not match.

05

Where the numbers come from

Every figure on these pages comes from a checked-in result document in the repository’s benchmarks directory or from the other product’s own documentation, dated where prices and feature lists can drift. The benchmark index lists every published study, and each result links its protocol, costs, and limits.

06

Sources