EN
Language · same page NLNederlands/mcp-geheugenservers-vergeleken/ ENEnglish (UK)/en/mcp-memory-servers-compared/ ESEspañolnot yet translated We do not remember your choice and never redirect you automatically.
About this choice

We only save your display preference in a preference cookie, for up to 180 days on this device. No CRM or advertising data. Your tracking consent stays unchanged. If cookies are blocked, your choice may not be saved.

DOC.K · KNOWLEDGE BASE

MCP memory servers compared: what seven servers do and do not do

This is a comparison of seven MCP memory servers, Mem0, Zep, Supermemory, Memlo, Membase, Mnemoverse and Jarvis, on ten fixed criteria: from how you connect and which AIs take part to price, data location, export and deletion. It is for anyone looking for a memory that more than one AI can read, whether you are building an agent yourself or using ChatGPT, Claude and Codex side by side.

Everything comes from the official documentation or product page of each product, read on 12 November 2026. Jarvis is our own product; we say so plainly, along with the points on which the others are further ahead. What we could not confirm, we label as such. We did not measure speed or accuracy.

Download the decision aid as a skill ↓3.5 KB · zip with SKILL.md · no account needed · works with Claude or Codex to ask the six questions up front, decide the type of memory and hold your candidates against the ten criteria in this article
CLAUDE CODE
  1. Unzip into ~/.claude/skills/ (or .claude/skills/ in your project).
  2. Claude then recognises the skill whenever you ask which memory server or shared memory suits your AIs.
  3. You can also call it directly with /geheugenserver-kiezen-en.
CLAUDE.AI
  1. Go to Customize, then Skills.
  2. Upload the zip there as a skill.
  3. Or paste the contents of SKILL.md into the project instructions of a Project.
CODEX
  1. Open AGENTS.md in your repo.
  2. Paste the contents of SKILL.md into it, or keep SKILL.md next to it and point to it from AGENTS.md.
  3. Codex reads that at the start of every session.
Illustration: seven frosted-glass discs in a row on a titanium rail on dark basalt, one of them lit up.
image · visual impression, not a photo or screenshot
DOC.01

What an MCP memory server is

An AI assistant starts every conversation with an empty head, unless it has a memory. The built-in memory of ChatGPT or Claude solves that for that one assistant. Use two or three side by side, and one still does not know what you told the other.

An MCP memory server is a memory that sits outside the assistant. It runs as a server at an address, and the assistant talks to it over the Model Context Protocol, MCP for short. The makers of MCP call it an open standard for connecting AI applications to external systems, and compare it to a USB-C port (modelcontextprotocol.io, read on 12 November 2026).

In practice your assistant gains a few tools: save something, search for something, and with some of the products delete something too. At Mem0 these are called add_memory and search_memories, for example (Mem0 documentation), and at Mnemoverse memory_write and memory_read (README on GitHub). The assistant decides for itself when to use them, or you set out in its instructions when it should.

There are two ways to connect. With most products you add an address in your assistant's settings and then sign in in the browser. That is called OAuth: you give permission without copying a key. The other way is an API key that you put in a configuration file. A few products can also be run locally, on your own machine.

Two things cannot be taken for granted. The first: that the memory is shared between assistants from different makers. Most products in this comparison can do that, but not all in the same way. The second: that the assistants can also coordinate with each other. A shared memory is a shared cupboard. Sending a message or handing over a task is something else, and we found that in only one of the seven.

Section 01 · The comparison

Seven memory servers, ten criteria

reg. G.001

Each column is a product, each row a criterion. Everything comes from the official pages of the product itself, read on 12 November 2026. Prices are in dollars, as the makers publish them. On a phone, swipe the table sideways.

FIG.01: MCP memory servers side by side, as of 12 November 2026sheet 1/2
Mem0ZepSupermemoryMemloMembaseMnemoverseJarvis (TheSEO)
Who it is forDevelopers and companies building their own AI appTeams building agents in production, mainly larger organisationsDevelopers; through MCP also teams using off-the-shelf assistantsPeople and teams with several agents; invitation only for nowBoth: users of AI apps and developersDevelopers and users of coding agentsAnyone using several off-the-shelf AIs side by side
How you connectSDK (Python, Node), API and CLI. MCP at mcp.mem0.ai/mcp, with browser sign-in or an API keySDKs (Python, TypeScript, Go). MCP at api.getzep.com/mcp, with OAuth through your organisation's sign-inAPI, SDKs (TypeScript, Python) and plugins. MCP at mcp.supermemory.ai/mcp with OAuth. Locally with a single binaryMCP only, at memlo.ai/mcp. OAuth in the browser, no API keyMCP at api.app.membase.com/mcp-http, with OAuth or a developer key. Also REST, SDKs, a Chrome extension and a local modeMCP at mcp.mnemoverse.com/mcp, with OAuth or an API key. Also a local package via npx, a Python SDK and RESTMCP only, at mcp.theseo.nl/mcp. Sign in through the browser (OAuth). No SDK
AIs mentioned by nameClaude, Claude Code, Codex, Cursor, Windsurf, VS Code, OpenCodeClaude, ChatGPT, Claude Code, Codex, Cursor. The plugin requires business plansChatGPT, Claude, Cursor, Claude Code, Codex, OpenCode, OpenClaw, HermesClaude Code, Claude, Codex, ChatGPT, GeminiChatGPT, Claude, Claude Code, Cursor, Codex, Grok, Kimi CodeClaude Code, Claude, Cursor, VS Code, ChatGPT, Gemini CLI, Zed, Cline, JetBrainsClaude, Claude Code, Codex, Cursor, ChatGPT on business plans
Shares memory between AIs from different makersThrough your account: memories live in your Mem0 account, not on your computer. The documentation does not promise it in so many wordsYes. One graph per user for in-house agents and MCP clientsYes. One shared memory layer for every assistant that speaks MCPYes. All connected agents read and write the same memoryYes, for reading. Writing is only possible with a developer key, not through the OAuth connectionYes, plus shared rooms for several agentsYes. The same account for all your AIs
Messages or tasks between AIsNot found in the documentation on 12 November 2026Not found in the documentation on 12 November 2026Not found in the documentation on 12 November 2026Not found in the documentation on 12 November 2026. It does have a work journal per agentNot found in the documentation on 12 November 2026Not found in the documentation on 12 November 2026. Its own documentation calls that a different kind of protocolYes. An Inbox for messages between AIs, Tasks and a Log
Self-hosted or hostedBoth. Hosted platform, or self-hosted with the open-source version. The MCP server belongs to the platformHosted, or in your own cloud on Enterprise. The old Community Edition is no longer supportedBoth. Cloud, your own cloud, or local. Local is free up to 10,000 documents, without the MCP serverHosted only. Self-hosting: not foundHosted, plus a local mode with your own OpenAI keyHosted. Self-hosting the engine is possible on Enterprise only, by agreementHosted only
Open sourceYes, Apache 2.0Not Zep itself. Graphiti, the framework underneath, is Apache 2.0Repository under MIT. According to the documentation, the server binary for self-hosting comes from a non-public codebaseNot foundSDKs, CLI and MCP server under MIT. The memory engine is closedMCP server and Python SDK under MIT. The engine is a hosted serviceNo
Pricing model on 12 November 2026Free: no charge, with your own LLM key 50,000 adds and 10,000 retrievals a month. Custom: on requestStart free with 10,000 credits a month. Flex $125 a month, Flex Plus $375 a month. Enterprise on requestFree: $0 with $5 of credits a month. Pro $19, Max $100, Scale $399 a month. Enterprise on requestNot published. Invitation onlyNot published. A free account has an initial allowance of chat turnsFree: $0 with 1,000 queries a day. Pro $29 a month, Team $149 a month. Enterprise on requestStart €9 and Solo €29 a month including VAT, can be ordered online. Team and Business not yet. You pay for your brain, not per AI query
Where the data is storedWith hosting providers in the US and other countries; the region is not given. Self-hosted: with youRegion not found in the documentation. On Enterprise in your own cloudNo region named. Database with Timescale, Cloudflare in front. Local: on your own machineDigitalOcean, San Francisco region (US)Amazon Web Services in Singapore. Local mode: on your own machineMemory API, database and MCP server in the Netherlands (Railway). Accounts in the USMessages and memory stay within one account. We found no region on our own pages
Export and deletionExport through the API. Deletion through the API and MCP. Account deletion on requestDeletion through the API and SDK. Export: not foundDeletion through MCP and the API, account on request. Export on request only, no button foundAn "Export my data" button. Deletion with 30 days to restore. The whole account can be deletedExport via Settings. "Delete everything" in SettingsNo export feature; on request by email. Individual memories are deleted through the REST API, the account by emailExport and deletion on the account page

Sources, all read on 12 November 2026: mem0.ai, Mem0 MCP documentation, Mem0 pricing, Mem0 on GitHub, Mem0 privacy policy; getzep.com, Zep Context MCP Server, Zep pricing, Zep FAQ, Graphiti on GitHub; supermemory.ai, Supermemory MCP documentation, Supermemory pricing, Supermemory self-hosting, Supermemory privacy policy; memlo.ai, Memlo for AI assistants, Memlo privacy policy; membase.com, Membase MCP documentation, Membase access control, Membase privacy policy; Mnemoverse on GitHub, Mnemoverse pricing, Mnemoverse privacy policy. The Jarvis column describes our own product, at the address of mcp.theseo.nl.

Section 02 · The choice

Choose this if...

reg. G.002

The table says what each product has. These cards say which situation it suits. There is no ranking in them: the order is that of the table.

FIG.02: Per product: when it is the logical choicesheet 2/2
01 · Mem0

Choose Mem0 if you are building an app yourself and want open source

You are building your own AI application, you want an SDK in Python or Node, and you want to keep the option of hosting it yourself later. The code is under Apache 2.0. Check the privacy policy for the free plan before you put client data into it.

02 · Zep

Choose Zep if your organisation runs agents in production

You have a team, an identity provider and requirements for access control. Zep lets employees reach the same memory with their own Claude or ChatGPT as the agents you built yourselves. For one person with two assistants it is too heavy.

03 · Supermemory

Choose Supermemory if you want to start quickly and also be able to run locally

One address, sign in with OAuth, and your assistants share a memory. The prices are simply on the site. If you want everything on your own machine there is a local version, but it does not have the MCP server or the connectors.

04 · Memlo

Choose Memlo if you want a simple shared memory and can wait

Memlo is made for people, not only for developers: version history, restore, teams with roles. It is invitation only, there is no price, and the maker itself writes that at this stage you should not use it as your only copy.

05 · Membase

Choose Membase if your memory also needs to come from your browser

Membase has a web app and a Chrome extension alongside MCP and SDKs, and you decide per app which memory it may read. An assistant connected through OAuth can only read. The hosted data is stored in Singapore.

06 · Mnemoverse

Choose Mnemoverse if you want your data in the Netherlands and need shared rooms

According to the privacy policy the memory data sits with a host in the Netherlands, the free plan is on the pricing page with its limits, and rooms let several agents work from one memory. Deletion and export do not yet come as a button.

07 · Jarvis (our own product)

Choose Jarvis if your AIs also need to keep each other informed

Jarvis is the only one of the seven where we found an inbox between AIs, tasks and a log. There is no SDK, you cannot host it yourself and it is not open source. Start and Solo can be ordered online; Team and Business are still in development.

Source: the same pages as under the table, read on 12 November 2026. The card on Jarvis describes our own product.

DOC.02

The three types: memory layer, memory that travels with you, and team space

The seven products look alike, but they do not solve the same problem. Make that distinction first, and you are left choosing between only two or three products.

1. The memory layer for developers. You are building an AI application yourself and want it to remember its users. You then need an SDK, an API, and control over where it runs. This is where Mem0, Zep and Supermemory come from. Mem0 calls itself the memory layer for AI agents and applications (mem0.ai), Zep the context layer for reliable agents (getzep.com). For these three the MCP server is an extra way in, not the core of the product.

2. Memory that travels with you between apps. You are not building anything. You use Claude, ChatGPT and a coding agent, and you want what you tell one to be known to the other. Memlo, Membase and Mnemoverse are made for this, and the MCP side of Supermemory does the same. Mnemoverse puts it like this: what you learned in Claude Code is there in Cursor an hour later (README).

3. The team space. Here it is not only about what the AIs know, but also about what they tell each other. Alongside a memory there is a mailbox, a task list and a log. That is where Jarvis sits. We have set out exactly what a team space is, and how it differs from the other two, in what is an AI team space.

The boundaries are not sharp. Zep lets employees' own assistants reach the same memory as the agents the company built, and so moves from type 1 towards type 2. Mnemoverse and Memlo have shared spaces for teams, and so touch on type 3. But a shared space with only a memory is not yet a team space: the AIs can read the same things there, not ask each other anything.

If type 2 is all you need, a team space is too much. How to fill and maintain one memory for three assistants is set out in one memory for ChatGPT, Claude and Codex.

DOC.03

How we compared: only what the makers themselves write

This comparison is desk reading, not a test. Nothing was run side by side. For each product we read the official sources and asked the same ten questions. That is a limitation, and it belongs at the top.

Which sources. For each product, the maker's own product page, documentation, pricing page, privacy policy and GitHub repository. No blogs, no comparison sites, and not the pages on which the makers compare themselves with each other either. There are plenty of those: Zep has a Mem0 alternative page and Supermemory a comparison with Mem0 (links on getzep.com and supermemory.ai). A maker is not a reliable source on its competitor.

Which date. Everything was read on 12 November 2026. This is a market in which prices and features change from month to month. Two of the seven products are still at an early stage, and a third has discontinued an earlier self-hosted version. So read any amount or limit in this article as the position on that day.

Which criteria. Who it is for, how you connect, which AIs are mentioned by name, whether the memory is shared between AIs from different makers, whether AIs can send each other messages or hand over tasks, self-hosted or hosted, open source and under which licence, the pricing model, where the data is stored, and export and deletion.

What "not found" means. Where the table says something was not found in the documentation, it means exactly that. It may exist without being described, or be on a page we did not read. It does not mean the maker cannot do it.

Our own product. Jarvis belongs to TheSEO. We held it against the same ten criteria and left in the answers that come out badly: no SDK, cannot be self-hosted, not open source, no region named for the data. You may expect us to know more about Jarvis than about the rest. You may not expect us to be impartial about it. That is why every other product has its source next to it, so you can check it yourself.

DOC.04

Mem0: memory layer for developers, open source and with a hosted MCP server

Mem0 is built for people making an AI application themselves. You install the SDK for Python or Node, and your app can store and retrieve memories per user (mem0.ai). The code is on GitHub under the Apache 2.0 licence, with three ways to use it: as a library, as a self-hosted server via Docker, or as a hosted platform (GitHub).

What it does well. Freedom of choice. You start on the hosted platform and can move to your own servers later, or the other way round. The MCP server is hosted at mcp.mem0.ai/mcp and, according to the documentation, works with Claude, Claude Code, Codex, Cursor, Windsurf, VS Code and OpenCode. Most clients open a browser window for you to sign in; where that is not possible, you use an API key. There are tools for deletion, one memory at a time or everything at once (MCP documentation).

Where it stops. The MCP server belongs to the platform, and the open-source version lacks some of the features. According to Mem0's own comparison, graph memory, webhooks and memory export, among others, are available on the platform only (platform versus open source). We found no messages or tasks between AIs. And the documentation does not say in so many words that two assistants share the same memory; it does say that memories live in your Mem0 account and not on your computer.

Price. There are two plans. Free costs nothing: for the first thirty days 10,000 adds and 1,000 retrievals a month with Mem0 covering the model costs, and with your own LLM key 50,000 adds and 10,000 retrievals a month. Above that you pay by usage, with a monthly cap you set yourself. Custom is on request, with no amount on the page (pricing page, 12 November 2026).

Data. The privacy policy says data is held with hosting providers, partly in the United States and other countries, without naming a region. There is one sentence you need to know: Mem0 uses certain information from the interactions of free-plan users to train its models, and does not do so with data from paying customers (privacy policy, version of 22 August 2026). Export goes through the API, with a schema you supply yourself (memory export).

DOC.05

Zep: context layer for organisations, with access control and its own graph

Zep is aimed at teams with agents in production. The site calls it the context layer for reliable agents and says agents connect through MCP or through the SDKs (getzep.com). Those SDKs exist for Python, TypeScript and Go. Underneath Zep sits Graphiti, an open-source framework for knowledge graphs under the Apache 2.0 licence (Graphiti on GitHub).

What it does well. Access control. The Context MCP Server runs at api.getzep.com/mcp and lets an employee reach their own memory in Zep with their own Claude, ChatGPT, Claude Code, Codex or Cursor. They sign in through the organisation's identity provider: Google Workspace by default, and other providers too on Enterprise. On sharing, the documentation is clear: the agent the company built and the user's personal MCP client work on the same graph (Context MCP Server).

Where it stops. Zep is not for one person with two assistants. An administrator has to switch the connection on first, and the plugin requires a Team or Enterprise plan from Claude, or ChatGPT Business, Enterprise or Edu (Zep in Claude and ChatGPT). Hosting Zep yourself on your own machine is no longer possible: according to the FAQ, the Community Edition has been discontinued. The alternatives are Zep Cloud, running Graphiti yourself, or Zep in your own cloud on Enterprise (FAQ). We found no messages or tasks between AIs.

Price. You start free with 10,000 credits a month and one seat on the MCP server. Flex costs $125 a month with 50,000 credits and five seats. Flex Plus costs $375 a month with 200,000 credits and fifteen seats. Enterprise is on request (pricing page, 12 November 2026). So you pay in credits, and the number of people who may connect through MCP depends on your plan.

Data. We found no region in the documentation. We did find three forms: Zep's cloud, Zep's cloud with your own encryption keys, or Zep inside your own network (security and compliance). Deletion is fine-grained through the API: individual facts, nodes, conversations (deleting data). We found no export feature.

DOC.06

Supermemory: one address with OAuth, clear prices and a local version

Supermemory says it is building the default engine for agent memory, available through an API, plugins and MCP (supermemory.ai). It started out as a product for developers, with SDKs for TypeScript and Python. The MCP documentation also addresses teams that do not write code.

What it does well. The barrier is low. You add mcp.supermemory.ai/mcp, sign in with OAuth, and you do not need an API key. The documentation says it plainly: every assistant that speaks MCP gets a shared memory layer. You divide the memory into spaces and decide, per connected app, what it may reach. There is a tool for forgetting something (MCP documentation). The site names ChatGPT, Claude, Cursor and Claude Code, and the repository adds plugins for Codex and OpenCode, among others (GitHub).

Where it stops. The local version is more limited than it sounds. You start it with one command and without a database, which is pleasant. But according to the documentation it is free up to 10,000 stored documents, it has no connectors and no Supermemory MCP, and the server binary comes from a codebase that is not public. The repository itself is under MIT (self-hosting). The product page calls the same binary open source (product page). Those two statements do not agree; we go by the documentation. We found no messages or tasks between AIs.

Price. Of the seven, this is the clearest pricing page. Free costs nothing and gives $5 of credits a month. Pro costs $19 a month, Max $100 and Scale $399, each with a monthly credit allowance that is higher than the price. Enterprise is on request (pricing page, 12 November 2026). You use up the credits by storing and searching.

Data. The privacy policy names no region. It names Timescale for the database and Cloudflare for delivery and security, and says data may be processed in other countries. Deletion of your account and data is available on request, and so is data portability, where technically feasible (privacy policy, updated 24 October 2026). We found no export button. The policy says nothing about training on customer data.

DOC.07

Memlo: shared memory for ordinary users, still invitation only

Memlo describes itself as memory for all your AI agents, shared over MCP (memlo.ai). The site has examples for developers, parents, executives and shopkeepers. That says something about the audience: this is not only for people with a terminal open.

What it does well. It is the simplest offer of the seven. There is one way in, memlo.ai/mcp, with MCP's standard OAuth and no API key. Every connected agent reads and writes the same memory. You can view every memory, correct it and restore it to an earlier version (how it works). There are teams with roles, from owner to viewer. And the page for AI assistants is precise about what does and does not work, with instructions per client for Claude Code, Claude, ChatGPT and Gemini (for AI assistants).

Where it stops. You cannot simply get in at the moment: accounts are by invitation, with a waiting list. There is no price. And the maker itself writes that the product is in early access, that data may change or disappear at this stage, and that you should not make Memlo the only copy of anything important. We found no SDK, self-hosting or open-source licence. No messages or tasks between AIs either. There is a work journal in which an agent can record what it did. That comes close to a log, but it is not a mailbox.

Price. Not published on 12 November 2026.

Data. Here Memlo is more precise than most. The service runs with DigitalOcean in the San Francisco region in the United States, with the backups in the same region. For ranking search results, queries and snippets go to Fireworks AI, for which Memlo gives no commitment that processing stays in the US (subprocessors). Memlo does not train on your content, unless you share it yourself as feedback. You export everything with a button in the web app, and a deleted memory can be retrieved for thirty days (privacy policy, version of 3 November 2026).

DOC.08

Membase: memory from your browser and your apps, with access per app

Membase calls itself the memory layer for AI agents and for the people who use them (membase.com). It comes from Unibase. It builds one long-term memory from your conversations, files and connected apps, and has a web app and a Chrome extension for that, alongside the ways in for developers.

What it does well. Breadth and control. You connect through MCP, REST or an SDK for Python or TypeScript. The documentation has a page of its own for ChatGPT, Claude, Claude Code, Cursor, Codex, Grok and Kimi Code (MCP documentation). For each memory you switch each app on or off, and according to the documentation that takes effect on the very next call. A permission expires after ninety days (access control). There is a local mode that keeps the memory on your own machine.

Where it stops. An assistant you connect through the consent screen can only read. Writing, deleting and forgetting are only possible with a developer key of the right level. That is a safe choice, but it means Claude in the browser will not add anything to your memory of its own accord, as it does with Memlo or Supermemory. The SDKs, the CLI and the MCP server are under MIT; the memory engine itself is closed (GitHub). The local mode requires Python 3.12 or 3.13 and your own OpenAI key. We found no messages or tasks between AIs; the site does mention sharing selected memories between agents.

Price. Not published. The FAQ says only that a free account has an initial allowance of chat turns, and that after that you need your own model source or a suitable plan (FAQ, 12 November 2026).

Data. The hosted service stores data with Amazon Web Services in Singapore. Model providers may process requests in other countries, including the United States. Membase says it does not use your content to train AI models. The site itself states that the hosted service is not end-to-end encrypted. You export in the settings, and that is also where you find a button to delete everything (privacy policy, updated 9 November 2026).

DOC.09

Mnemoverse: hosted memory with shared rooms and data in the Netherlands

Mnemoverse is a hosted long-term memory for AI agents over MCP. The MCP server is on GitHub under the MIT licence; the memory engine it talks to is a hosted service (README on GitHub). We mentioned it earlier in the article on the team space, because it has shared rooms.

What it does well. Three things. It connects widely: at mcp.mnemoverse.com/mcp with browser sign-in, or with an API key, with instructions for Claude Code, Claude, Cursor, VS Code, ChatGPT and Gemini CLI, among others. It has rooms: you create a room, invite someone else with a code, and several agents work from one memory. And it is the only one of the seven that names a host in the Netherlands for the memory data (privacy policy, updated 11 November 2026).

Where it stops. Rooms are shared memory, not coordination. Mnemoverse says so itself in its documentation: a task protocol can hand over work and carry messages, a memory keeps what agents will need later, and those are two different things (shared memory for multiple agents). Self-hosting the engine is possible on Enterprise only, by agreement. And deletion is not finished yet: you delete individual memories through the REST API, not through MCP or the dashboard. Deleting your account is done by email, within thirty days of verification. Deleting your account yourself is planned for 31 December 2026. There is no export feature; you can request a copy.

Price. Free costs nothing and needs no credit card: 1,000 queries a day, 10,000 atoms (the unit in which Mnemoverse counts storage) and 60 requests a minute. Pro costs $29 a month, Team $149 a month, and Enterprise is on request (pricing page, 12 November 2026).

Data. The memory API, the database and the MCP server run with Railway in the Netherlands. The accounts database is with Neon in the United States, and the embeddings are computed by OpenAI or NVIDIA in the United States. So your memory sits in the EU, but your text does travel to the US for processing. Mnemoverse says it does not use private memory to train public or general AI models.

DOC.10

Jarvis: our own product, a young team space

Jarvis belongs to TheSEO, and so to us. Read this part with that in mind. We keep to the same criteria and write down only what is established.

What it is. A team space with four parts: Memory, an Inbox for messages between AIs, Tasks and a Log. The address is mcp.theseo.nl/mcp. You sign in through the browser with OAuth, and you use the same account for all your AIs. It works with any AI that speaks MCP: Claude, Claude Code, Codex, Cursor, and ChatGPT on business plans. There is more about the product on the Jarvis page, and how to add the address is on our MCP page.

What it does well. The one thing we did not find in the other six: AIs letting each other know things. Our own experience dates from 11 November 2026. That day Claude Code and Codex coordinated a deploy with each other through the Inbox. How that went, and which rules came out of it, is set out in Claude Code and Codex working together. Messages and memory stay within one account. Export and deletion are on the account page.

Where it stops. At almost everything the others are strong in. There is no SDK, so you do not build your own app on it as you would with Mem0, Zep or Supermemory. You cannot host it yourself and it is not open source. Mem0 is, and Supermemory and Membase have a local version. We found no region for the data on our own pages, whereas Memlo, Membase and Mnemoverse do name one. And our own evidence is small: two AIs and one coordinated deploy.

Price. On 12 November 2026 Start costs €9 and Solo €29 a month, including VAT for personal use. Both can be ordered online with a TheSEO account, for customers in the Netherlands (the Jarvis page). Team and Business are still in development and cannot be ordered. You pay for your brain and not per AI query; your own AI subscription stays separate.

What has been added since 12 November. Live since 12 November 2026, so after the table was compiled: messages to a single session instead of to a whole AI, waiting for a message, threads with a type per message, and reserving a part that an AI is working on. We have not counted it above; the table gives the state as of 12 November during the day.

Why is Jarvis in a comparison of memory servers if it is a team space?

Because the memory is one of its four parts, and because people looking for a memory server often really want their AIs to know from each other what is going on. If you are only looking for a memory, the other six are the more logical choice and a team space is more than you need.

DOC.11

Privacy and data: where it is stored, who trains on it and how you remove it

A memory server holds exactly the things you do not want to explain again every time: how your business works, who your clients are, what you have decided. That makes the question of where it is stored more important than with most software. This is what the makers themselves write about it.

Where it is stored. Three of the seven name a place. Memlo: San Francisco in the United States. Membase: Singapore. Mnemoverse: the Netherlands for the memory, the United States for accounts and embeddings. Mem0 names the United States and other countries without a region. We found no region for Zep or Supermemory, nor for Jarvis. If you want to be certain where it is stored, that leaves the versions you run yourself: Mem0 open source, Supermemory local, Membase local, or Zep in your own cloud on Enterprise.

Who trains on it. Here the answers diverge. Mem0 trains on interactions from the free plan and not on those of paying customers. Memlo does not train, unless you share something as feedback. Membase says it does not train on your content. Mnemoverse says it does not use private memory for public or general models. We found no statement on it in Supermemory's privacy policy or in Zep's security documentation. The sources are in the sections on each product above.

Export. Memlo, Membase and Jarvis have a button to download everything. Mem0 has an export through the API. With Supermemory and Mnemoverse you request a copy. We found no export feature for Zep.

Deletion. All seven let you delete something, but not equally easily. Memlo, Membase and Jarvis have it in the app or on the account page. Mem0, Zep and Supermemory do it through the API or an MCP tool. With Mnemoverse an individual item goes through the REST API and your account by email. Check what deletion means, too. With Memlo a deleted memory stays recoverable for thirty days. With Membase, copies in backups expire within 35 days. With Supermemory, forgetting takes a memory out of the search results while the record stays in the database (product page).

What you can do yourself. Do not put passwords, keys or tokens in a memory, with any of the seven. Anything an AI can read, it can repeat somewhere else. Do not put client data in it either before you know where it is stored and whether you need a data processing agreement. For a Dutch business with client data, the region is not a detail. We are not lawyers; put it to someone who is before you fill a service outside the EU with personal data.

DOC.12

What we deliberately did not assess

A comparison is also what is left out. These are the things you may have expected that are missing, with the reason.

Speed. Not measured. How fast a memory answers depends on your region, your plan, the load and the size of your memory. Supermemory quotes a median on its pricing page and Zep a threshold in the comparison with Graphiti. Those are figures from the maker about itself, under conditions we do not know. So we do not repeat them.

Retrieval accuracy. Not measured. This is probably the most important criterion, and the only one you cannot get from documentation. Does the memory find the right fact, even when you phrase it differently? Does it forget what is out of date? Mem0, Zep, Supermemory and Membase each have a research or benchmark page of their own for it, and Memlo cites a saving in tokens compared with memory in text files. Those are the maker's measurements of its own product. We have not checked the sums and do not repeat them. A fair measurement needs the same set of facts and questions on all seven, and we have not done that.

Quality of what is remembered. The products differ in how they turn a conversation into a memory: literally what the AI supplies, or a model of their own that extracts facts and links them together. That difference is large, and you only see it after weeks of use. Not assessed.

Certificates. Mem0, Zep and Supermemory list certifications such as SOC 2 on their sites. We have not requested or checked any reports. If you need them, ask the maker for them.

Reliability and support. Not tested. It is relevant, though: Memlo itself says it is at an early stage, Jarvis is young, and Zep has discontinued an earlier self-hosted version. In this market, today's product will be something else in a year, or gone. That is why export and deletion weigh more heavily in our choice than an extra feature.

If you want to measure for yourself, keep it small. Put the same twenty facts into two products, ask the same ten questions a week later, and see which answers are right. That tells you more than any table, this one included.

DOC.13

How to choose: five questions in this order

The decision aid you can download at the top of this article takes you through these questions and keeps to the table above. In short:

  • Are you building something, or using something? If you are building your own app, look at Mem0, Zep or Supermemory. If you use off-the-shelf assistants, look at Memlo, Membase, Mnemoverse, Supermemory or Jarvis.
  • Do your AIs only need to know the same things, or also to coordinate with each other? Knowing only: a memory is enough. Coordinating, such as announcing a deploy or handing over a task: then you need a team space, and in this comparison that is only Jarvis.
  • Where may the data be stored? If it has to be in the EU or with you, a few drop out straight away. Look at the row on data and at the section on privacy.
  • What must be possible if you want to leave? An export button and a delete button. Check them before you put anything in, not after.
  • What does it cost at your level of use? Work in the maker's unit: credits at Zep, credit allowance at Supermemory, queries a day at Mnemoverse, adds and retrievals at Mem0. If there is no price, you do not yet know what it costs.

Start with the first two questions. They bring the list of seven down to two or three. There is more background on MCP, memory and AIs working together in the knowledge base.

DOC.16 · FAQ

Frequently asked questions about MCP memory servers

What is an MCP memory server?

A memory that sits outside your AI assistant and that the assistant can reach over the Model Context Protocol. It gains tools to save something, search for it and usually delete it too. Because the memory is separate from the assistant, with most products several AIs can use the same memory.

Which MCP memory server can you host yourself?

On 12 November 2026: Mem0 in the open-source version, although the MCP server there belongs to the hosted platform. Supermemory has a local version, free up to 10,000 documents and without the MCP server. Membase has a local mode with your own OpenAI key. Zep (in your own cloud) and Mnemoverse (by agreement) offer it on Enterprise only. We did not find it for Memlo, and Jarvis cannot be self-hosted.

Which one is open source?

Mem0 is under Apache 2.0. With Zep, the underlying framework Graphiti is open source under Apache 2.0; Zep itself is not. With Supermemory, Membase and Mnemoverse the repository, the SDKs or the MCP server are under MIT, but the engine or the server binary is closed. We found no licence for Memlo. Jarvis is not open source.

Can AIs send each other messages or hand over tasks through a memory server?

For six of the seven we did not find that in the documentation on 12 November 2026. They share a memory, not a mailbox. Mnemoverse and Memlo have shared rooms or teams, but there too it is about memory. Jarvis has an Inbox for messages between AIs, Tasks and a Log; it is our own product.

What does an MCP memory server cost?

Four of the seven publish a free plan: Mem0, Zep, Supermemory and Mnemoverse. On 12 November 2026, paid plans start at $19 a month at Supermemory, $29 at Mnemoverse and $125 at Zep. Besides the free plan, Mem0 only has a plan on request. Memlo and Membase publish no price. Jarvis starts at €9 a month for the Start plan.

Which memory server keeps data in the EU?

Only Mnemoverse names a host in the Netherlands for the memory, the database and the MCP server; accounts and embeddings there do go to the United States. Memlo names San Francisco, Membase Singapore, Mem0 the United States and other countries. We found no region for Zep, Supermemory or Jarvis. If you want certainty, choose a version you run yourself.

Does one memory work with ChatGPT and Claude at the same time?

Yes, with most. Zep, Supermemory, Memlo, Membase, Mnemoverse and Jarvis name both ChatGPT and Claude. Mem0's MCP documentation names Claude, Claude Code, Codex and Cursor, but not ChatGPT. Mind the conditions: Zep's plugin requires a business plan for ChatGPT, Jarvis works with ChatGPT on business plans, and with Membase an assistant connected through the consent screen can only read.

Is Jarvis better than the other six?

No, it is different. For an app of your own a memory layer with an SDK is better, for self-hosting or open source too, and six of the seven are further along as a product. Jarvis does one thing we did not find in the rest: letting AIs coordinate through messages, tasks and a log. If you do not need that, you do not need Jarvis.

DOC.K · Next step

Choose the type first, then the product

Most of the doubt disappears once you know whether you are looking for a memory layer, a memory that travels with you or a team space. After that, data location, export and price decide. The decision aid at the top of this article asks those questions in the right order.

If you want to see what the team space looks like with us, including the messages between Claude Code and Codex, book a call. If you only want a memory, the table is all you need.

Written by GianlucaFounder. Has been building websites and search visibility for Dutch businesses since 2017, and has had his own AIs coordinate with each other through Jarvis since 11 November 2026. More about the institute.

Updated on 12 November 2026