MemoryRouter for your AI tools

One memory for every AI you connect.

Connect MemoryRouter to Claude, ChatGPT, Cowork, Claude Code, Codex, OpenClaw, and compatible MCP clients. Choose a vault and permissions, then let each supported AI search or save memory according to its actual integration.

Choose your AITransfer what ChatGPT can currently recall
Preparing a complete history import?

Connection enables future memory access. It does not read or import your old conversations.

You approve a durable memory

Your MemoryRouter vaultOne selected vault
ClaudeModel-directed tools
ChatGPTModel-directed tools
Coding agentsAutomatic hooks where installed
Same vault. Different host behavior.

Your first cross-AI win

Carry one approved memory from ChatGPT into Claude.

Connect both model-directed hosts to the same vault, move only what ChatGPT can currently recall/expose, then prove Claude can continue from it.

First-run successClaude recalls one reviewed detail from ChatGPT through MemoryRouter and continues the work.

  1. Connect MemoryRouter in ChatGPT.

    Search and save portable memory from an eligible Developer mode conversation. This is a manual Developer mode connection, not a Plugin directory install.

    Check: Ask ChatGPT to use MemoryRouter memory_status. A visible tool result proves the connector was used.

    Important: Opens the manual guide. This is not a Plugin directory install.

    View ChatGPT setup

  2. Transfer what ChatGPT can currently recall/expose.

    Invoke the reviewed transfer workflow, inspect every fact and inference, remove anything sensitive or uncertain, then deliberately save only the version you approve.

    Check: Search MemoryRouter for one approved profile detail before moving to Claude.

    Important: ChatGPT Skills can guide this workflow when invoked directly or when the request matches. They do not run passively, expose full history, or capture every conversation.

    Open the reviewed transfer workflow

  3. Review and connect MemoryRouter in Claude.

    Use one selected MemoryRouter vault from Claude web, desktop chat, or mobile chat. Choose the same MemoryRouter vault used in ChatGPT so the approved memory has one destination.

    Check: Claude shows MemoryRouter as available for the conversation and can run memory_status.

    Important: Opens a prefilled review. You still sign in, choose a vault, and approve access.

    Review and connect

  4. Prove continuation in Claude.

    Ask Claude to use MemoryRouter to recall one approved detail from the transfer, then continue the work from that detail. This explicit recall proves cross-AI continuity.

    Check: Claude returns the approved detail through a visible MemoryRouter tool result and uses it in the next response.

    Important: ChatGPT and Claude web connectors are model-directed. Neither connector passively captures every conversation or guarantees memory use on every turn.

    See the verification sequence

This is cross-AI continuity, not transcript sync. ChatGPT and Claude decide when to call MCP tools. Neither host passively sends every conversation to MemoryRouter.

Need deterministic history recovery? Keep that separate and prepare your official export.

Choose your AI

Start where you already work.

Each platform reaches the same MemoryRouter idea through a different setup. The card tells you whether memory is automatic, model-directed, or client-dependent before you connect.

Guided connections

Start in a familiar hosted app. These tools decide when to call memory, so explicit requests and verification matter.

Claude

Available: guided setup

Use one selected MemoryRouter vault from Claude web, desktop chat, or mobile chat.

Best for
People who want portable memory in hosted Claude chat.
Setup
Custom connector + OAuth
Difficulty
Easy
Required plan/app
Claude account; Free supports one custom connector
Memory behavior
Model-directed tools, not guaranteed every turnOne explicit MCP tool call chosen by Claude; writes require an explicit user request or confirmation.
Permissions
OAuthReadWriteMemoryRouter OAuth. No Memory Key is entered in the host.
Old conversations
Connecting does not import earlier Claude conversations or files.
Review and connect

Opens a prefilled review. You still sign in, choose a vault, and approve access.

Claude setup details

Prerequisites

  • Claude account; Free supports one custom connector
  • MemoryRouter account with at least one vault
  • Organization policy may require an Owner or Primary Owner to add/allow the connector

Verify

  • Confirm Connected in Customize > Connectors
  • In a new chat, explicitly ask Claude to run memory_status

Upgrade and remove

  • Hosted connector uses the live service; refresh or reconnect after material metadata changes
  • Disconnect MemoryRouter in Claude
  • Remove the custom connector if desired; stored vault memories are not deleted

Limits

  • Connector must be enabled for the chat
  • No deterministic hosted-chat lifecycle hook is installed
  • Store acknowledgement is queued, not an exactly-once durable item receipt

ChatGPT

Available: manual Developer mode

Search and save portable memory from an eligible Developer mode conversation.

Best for
Advanced ChatGPT users evaluating a private connector before directory launch.
Setup
Manual Developer mode
Difficulty
Moderate
Required plan/app
Eligible ChatGPT Pro, Plus, Business, Enterprise, or Education web account
Memory behavior
Model-directed tools, not guaranteed every turnOne model-selected MCP search or one explicitly requested/confirmed concise store call.
Permissions
OAuthReadWriteMemoryRouter OAuth. No Memory Key is entered in the host.
Old conversations
Connection does not import ChatGPT history. A deterministic archive importer is not shipped.
Host directory
not listed
View ChatGPT setup

Opens the manual guide. This is not a Plugin directory install.

ChatGPT setup details

Prerequisites

  • Eligible ChatGPT Pro, Plus, Business, Enterprise, or Education web account
  • Developer mode enabled in Settings under Security and login
  • Workspace policy permits Developer mode where managed

Verify

  • Enable MemoryRouter in a new conversation
  • Explicitly call memory_status and a read test before enabling writes

Upgrade and remove

  • Refresh/recreate the developer connection after metadata changes
  • Disable or remove the developer connection; vault data remains

Limits

  • Not automatic every turn
  • Developer mode/plan/workspace tool availability varies
  • Write returns queued acceptance with no stable item ID

Claude Cowork

Available: guided setup

Use MemoryRouter tools in Cowork tasks and optionally add behavioral skills/plugins.

Best for
Paid-plan Cowork users who want guided recall and deliberate saves.
Setup
Custom connector; optional package
Difficulty
Easy
Required plan/app
Paid Claude plan with Cowork access
Memory behavior
Model-directed tools, not guaranteed every turnOne model-selected search or one explicitly requested/approved durable-memory write, up to 12,000 characters.
Permissions
OAuthReadWriteMemoryRouter OAuth. No Memory Key is entered in the host.
Old conversations
Cowork may review an extracted export and save selected summaries, but no deterministic archive importer exists.
Host directory
not listed
Open Cowork setup

Opens the guide. You still add the connector, complete OAuth, and enable it for the task.

Claude Cowork setup details

Prerequisites

  • Paid Claude plan with Cowork access
  • MemoryRouter account with at least one vault
  • Organization policy can require Owner or Primary Owner enablement

Verify

  • Call memory_status and verify the selected opaque vault reference
  • Test an explicit recall and an explicitly approved save

Upgrade and remove

  • Reconnect after scope changes
  • Replace the optional ZIP with a newer validated version
  • Disconnect/remove the connector and optional artifact; vault data remains

Limits

  • No package hook currently guarantees recall or capture
  • Single writes are queued with no stable item ID
  • No archive parser/checkpoint/resume/reconciliation

Automatic local integrations

Install host lifecycle hooks for automatic recall and capture according to each package contract.

Claude Code

Available: published package

Automatic project recall and capture through local lifecycle hooks, plus optional MCP tools.

Best for
Developers who want cross-session project memory.
Setup
Published npm hooks package
Difficulty
Technical
Required plan/app
Node.js 18+
Memory behavior
Automatic hooks where installedPublished v2.0.0 captures a broader meaningful transcript delta with byte cursors and pending-batch retry; it does not yet prove exactly one prompt plus one final answer.
Permissions
ReadWriteA local Memory Key powers hooks; optional MCP uses OAuth.
Old conversations
The package captures future Claude Code work; it does not import prior Claude chat history.
Install for Claude Code

Opens the versioned install guide. Automatic hooks require the local package and Memory Key.

Claude Code setup details

Prerequisites

  • Node.js 18+
  • Claude Code installed and signed in
  • MemoryRouter Memory Key for automatic hooks

Verify

  • npx -y [email protected] doctor
  • Start a new same-project session and verify recalled context

Upgrade and remove

Limits

  • At-least-once local retry behavior; public v2.0.0 has no server event key or end-to-end exactly-once guarantee
  • Published transcript-delta semantics are broader than the unreleased typed prompt/final recovery target
  • Hook failures fail open

OpenAI Codex

Available: published package

Automatic project recall/capture with local hooks and explicit MCP tools.

Best for
Codex CLI, IDE, and local-host users in trusted projects.
Setup
Published npm hooks package
Difficulty
Technical
Required plan/app
Node.js 18+
Memory behavior
Automatic hooks where installedPrompt retrieval; completed turn with stable tool inputs/results and final answer; stable subagent outcome; compaction/session checkpoints.
Permissions
ReadWriteA local Memory Key powers hooks; remote MCP uses OAuth.
Old conversations
Local Codex can inspect files, but no supported deterministic ChatGPT archive importer exists.
Install for Codex

Opens the secure, versioned local install guide.

OpenAI Codex setup details

Prerequisites

  • Node.js 18+
  • Codex 0.145+ for full hooks; 0.146+ for current remote OAuth
  • MemoryRouter key for hooks

Verify

  • npx -y [email protected] doctor --scope user
  • Start a new trusted-project session and verify recall

Upgrade and remove

Limits

  • API key required separately from OAuth for hooks
  • No server-side exactly-once idempotency
  • Hosted tool details may be absent unless reflected in final output
  • Hooks depend on trusted local host support

OpenClaw

Available: published package

Automatic relay-based recall/capture plus optional historical workspace/session upload.

Best for
OpenClaw operators who want automatic agent memory.
Setup
Published npm plugin
Difficulty
Technical
Required plan/app
Current OpenClaw
Memory behavior
Automatic hooks where installedbefore_agent_start/LLM input recall and agent_end direct user/assistant conversation capture.
Permissions
A Memory Key stays in the local plugin configuration.
Old conversations
Explicitly scans configured workspace/session paths; separate from ongoing connection and capture.
Host directory
unverified
Install for OpenClaw

Opens the explicit npm-source install guide.

OpenClaw setup details

Prerequisites

  • Current OpenClaw
  • npm/network access
  • Explicit trust for the third-party npm source
  • MemoryRouter key

Verify

  • openclaw mr status
  • openclaw plugins inspect mr-memory --runtime --json

Upgrade and remove

  • openclaw plugins update mr-memory
  • openclaw mr off
  • openclaw plugins uninstall mr-memory

Limits

  • Tool use, subagent work, and internal processing are excluded
  • Relay mode does not proxy provider calls or provider API keys
  • Historical upload is a separate explicit command

Advanced connection

Bring another compatible MCP client and verify transport, OAuth, tools, and scope behavior before relying on it.

Other MCP clients

Advanced: client-specific

Connect any compatible remote MCP client to one MemoryRouter vault.

Best for
Advanced users whose client supports the required transport and OAuth behavior.
Setup
Client-specific remote MCP setup
Difficulty
Advanced
Required plan/app
Client supports remote Streamable HTTP MCP
Memory behavior
Client-dependent behavior, not guaranteed every turnClient-selected MCP calls against the shared tool catalog.
Permissions
OAuthReadWriteMemoryRouter OAuth. No Memory Key is entered in the host.
Old conversations
MCP connection alone never imports historical conversations.
Read generic MCP setup

Opens client-dependent instructions. Transport support alone does not prove compatibility.

Other MCP clients setup details

Prerequisites

  • Client supports remote Streamable HTTP MCP
  • Client supports RFC 9728 OAuth discovery and an OAuth callback flow
  • Client can expose the needed tools/resources and scopes

Verify

  • Run memory_status if the client exposes it
  • Test read before requesting write or delete scopes

Upgrade and remove

  • Refresh/reconnect according to the client after service metadata changes
  • Disconnect/remove the MCP server from the client; vault data remains

Limits

  • Compatibility must be verified per client
  • Transport support does not imply OAuth callback or tool UX support
  • No generic lifecycle capture guarantee

Three different jobs

Choose the kind of continuity you need.

Connection, a quick reviewed transfer, and a complete archive import are separate paths.

01

Going forward

Connect future memory

Authorize a supported AI to search or save the MemoryRouter vault you choose. Model-directed hosts may ask before tool calls. Hook integrations follow their documented lifecycle.

Available nowChoose your AI
02

A few minutes

Transfer what ChatGPT can currently recall/expose

Generate a structured profile, review every line, remove sensitive or uncertain items, then deliberately save the approved version. Useful, but not complete.

Available nowUse the reviewed prompt
03

Official archive

Prepare for deterministic import

Request and keep your official ChatGPT export. MemoryRouter's complete, receipt-backed importer is not available yet. Do not generic-prompt the archive.

Not available yetSee export and launch requirements

A visible trust boundary

Connect without handing over your account.

MCP is only the connector under the hood. You stay in control of the host, selected vault, and requested permissions.

  1. Pick a supported AI.

    Use the chooser to see the real setup, plan, and capture behavior.

    Check: The card matches the exact host and surface you intend to use.

  2. Review access.

    Sign in to MemoryRouter, choose one vault, and approve only the permissions the connection requests.

    Check: The host shows the expected MemoryRouter connection and selected vault boundary.

  3. Verify a real tool call.

    Ask for memory status, then explicitly test recall or an approved save. Connection status alone does not prove memory was used.

    Check: A visible MemoryRouter tool result appears in the conversation.

Prove the connection

A connected badge is only step one.

Follow the trail from setup to a visible MemoryRouter result. Never assume the host called memory just because authorization succeeded.

01

Open setup

Start from the platform card and read the exact boundary.

02

Review endpoint

https://mcp.memoryrouter.ai/mcp
03

Choose a vault

Select one vault during MemoryRouter authorization.

04

Review permissions

Treat read, write, and destructive access as separate capabilities.

05

Verify status

Ask the AI to run MemoryRouter status.

06

Verify across an AI

From another connection to the same vault, explicitly recall one approved fact.

A connected badge proves authorization. A visible MemoryRouter tool result proves use.

Privacy and permissions

You choose what crosses the boundary.

A normal MemoryRouter connection does not receive your ChatGPT history, browser cookies, email, password, or every file on your device. OAuth binds the connection to one MemoryRouter vault you select. The host can use only the tools and permissions granted to that connection. Disconnecting the host stops future access but does not delete memories already stored in the vault.

What MemoryRouter does not get

No account password. No browser cookies. No mailbox access. No hidden ChatGPT memory database. No automatic permission to old conversations. An archive import, when supported, uses the official file you choose and a separate approval flow.

RRead
Search and inspect memory in the selected vault.
WWrite
Save explicitly requested or approved durable memory. A write grant is separate from read where the host exposes independent scopes.
!Destructive controls
Require separate authorization and explicit confirmation where exposed. Use the MemoryRouter dashboard for individual memory review and curation.
FFiles
A connector does not gain general device access. Cowork or a local coding agent can read only files or folders the user grants through that host.
×Revocation
Disconnect the host or revoke the MemoryRouter authorization. Stored vault data remains until the user deletes it separately.

Connect versus import

Know which promise you are choosing.

Future tool access is not a history migration. A reviewed profile is not an official archive import.

Compare connecting future memory, quick transfer, and complete import
QuestionGoing forwardConnectAvailable nowA few minutesQuick transferAvailable nowOfficial archiveComplete importNot available yet
Main purposeFuture recall/savePortable profile nowDeterministic history recovery
SourceNew host tool calls or hook eventsContext ChatGPT can currently recall or expose to the modelOfficial OpenAI export selected by the user
Automatic?Depends on platformNo, the user runs and reviews a promptUnavailable until the deterministic importer ships
Complete?No historical importNoOnly within an explicit supported-content policy after release
Review before writeHost and tool policyRequiredRequired preview and approval digest after release
Resume and cancelNot applicableNot applicableRequired after release
ReceiptOrdinary tool response or statusVerify with searchReconciled immutable import receipt after release

Going forward

Connect

Give a supported host access to one selected MemoryRouter vault.

Available now

Main purpose
Future recall/save
Source
New host tool calls or hook events
Automatic?
Depends on platform
Complete?
No historical import
Review before write
Host and tool policy
Resume and cancel
Not applicable
Receipt
Ordinary tool response or status
Choose your AI

A few minutes

Quick transfer

Create and review a portable profile from context ChatGPT actually surfaces.

Available now

Main purpose
Portable profile now
Source
Context ChatGPT can currently recall or expose to the model
Automatic?
No, the user runs and reviews a prompt
Complete?
No
Review before write
Required
Resume and cancel
Not applicable
Receipt
Verify with search
Use the reviewed prompt

Official archive

Complete import

A deterministic archive path with preview, approval, resume, cancellation, and reconciliation.

Not available yet

Main purpose
Deterministic history recovery
Source
Official OpenAI export selected by the user
Automatic?
Unavailable until the deterministic importer ships
Complete?
Only within an explicit supported-content policy after release
Review before write
Required preview and approval digest after release
Resume and cancel
Required after release
Receipt
Reconciled immutable import receipt after release
Prepare your official export

Questions, answered plainly

What connection actually means.

No protocol lesson required. Just the boundaries that matter before you connect.

What is MCP in normal language?

MCP is a standard way for an AI app to use an outside tool. Here, the tool is MemoryRouter. The connection lets the AI search or save the vault you approve. You do not need to understand the protocol to use a supported setup.

Does connecting MemoryRouter copy all my old chats?

No. Connection enables future tool access. It does not unlock a hidden ChatGPT or Claude history feed and it does not import old conversations.

Can MemoryRouter read everything in my AI account?

No. The connector receives only the arguments a host sends to an approved MemoryRouter tool. It does not receive your password, cookies, email, or a bulk copy of private account history.

Do I need to code?

Not for the guided Claude, ChatGPT, or Cowork paths. Claude Code, Codex, OpenClaw, and generic MCP setups are more technical and list their prerequisites before you start.

Which option is easiest?

For a normal Claude user, start with the guided custom connector. For ChatGPT, use Developer mode only if your account is eligible. For local coding work, use the published hook package for your coding agent.

Does every platform save memory automatically?

No. Claude, ChatGPT, and the current Cowork package use model-directed tools. Claude Code, Codex, and OpenClaw have local or host integrations that can call MemoryRouter automatically according to their documented lifecycle.

Can I use read-only access?

Where the host and current MemoryRouter install method expose a read-only grant, yes. Read-only must be enforced by OAuth scope or server policy, not merely by instructions in a skill.

Can I disconnect it?

Yes. Disconnect or remove the connection in the host, and revoke authorization if needed. That stops future access. It does not delete memories already stored in MemoryRouter.

Is this the same as ChatGPT Memory?

No. ChatGPT Memory is an OpenAI product feature. MemoryRouter is a separate vault that supported connected tools can use. ChatGPT does not expose its complete hidden memory or conversation history to MemoryRouter through MCP.

How do I move what ChatGPT can currently recall about me?

Use the quick-transfer guide for a reviewed, model-generated profile based only on what ChatGPT can currently recall or expose. For full history, request the official ChatGPT export and follow the complete-import guide current availability notice. Do not use a generic prompt to import an entire archive.

Does MemoryRouter work across Claude and ChatGPT?

The same selected MemoryRouter vault can be connected to supported Claude and ChatGPT surfaces. Actual recall and saving still follow each host's permissions and model-directed behavior.

Start where you already work

Choose one AI. Keep the same vault ready for the next.

See the real setup, permissions, and memory behavior before you connect.

Choose your AITransfer what ChatGPT can currently recall