Platform Comparison

Canonical MemoryRouter setup, capture, distribution, and historical-import behavior for every supported consumer surface.

Every platform can reach a selected MemoryRouter vault, but the host determines how reliably memory is used. These records are reviewed against shared/platform-truth.json.

Automatic hooks run locally at named lifecycle events. They do not depend on the model deciding to call an MCP tool.
Model-directed tools run only when the hosted model chooses the connector call. They are useful, but not guaranteed every turn.
Client-dependent MCP means transport support alone is insufficient. OAuth, tools, resources, scopes, and invocation behavior must be verified for that client.

Choose a platform

Platform	Recommended path	Current capture	Historical import	Distribution status
Claude	Review and add the prefilled custom connector	Model-directed search and explicitly requested/approved single-memory writes	No	Manual connector live; no MemoryRouter marketplace listing
Claude Cowork	Custom connector, optional repository skill/plugin	Model-directed search and one approved write up to 12,000 characters	No supported archive import; guided curation only	Manual connector live; ZIP repository-only; no public marketplace listing
ChatGPT	Manual Developer mode MCP setup	Model-directed search and one approved concise write	No	Manual Developer mode available to eligible accounts; not submitted/listed in Plugin directory
Claude Code	`npx -y memoryrouter-claude init`	Automatic typed prompt/final hooks	No	`memoryrouter-claude` published on npm
Codex	Secure key on stdin, `memoryrouter-codex`, then OAuth	Automatic message-only user request + final response capture	No supported ChatGPT archive import; curated experiments only	`memoryrouter-codex` published on npm
OpenClaw	`openclaw plugins install npm:mr-memory`	Automatic relay recall plus direct user/assistant capture at agent end	Yes, separate `openclaw mr upload` command	`mr-memory` published on npm; ClawHub status unverified
Other MCP clients	Add `https://mcp.memoryrouter.ai/mcp`	Client-directed; verify per host	No	Client-dependent compatibility

Connection and historical import are independent. Adding an MCP connector never imports old conversations. OpenClaw is the only current consumer path with a first-class historical upload command.

Published memoryrouter-claude creates typed prompt/final events instead of sweeping the transcript. It captures one user prompt and one final user-visible assistant response per completed turn, requires a durable accepted/duplicate receipt before advancing local state, and excludes tool, intermediate, compaction, failure, and subagent transcript content. Hook failures fail open.

Codex

Published memoryrouter-codex 1.1 hooks retrieve on UserPromptSubmit; store only the user request and final assistant response for completed turns; capture stable final subagent outcomes separately; and make message-only checkpoints before compaction/session end. PostToolUse is not installed, and tool names, IDs, inputs, results, commands, and patches are excluded. The hook API key is separate from remote MCP OAuth, and no server-side exactly-once guarantee exists.

OpenClaw

The npm plugin uses relay hooks. Recall occurs before agent/model input and direct user/assistant conversation is captured at agent end. Tool use, subagent work, and internal processing are excluded. Inference and provider API keys remain in OpenClaw; this plugin is not a provider-key proxy.

Hosted Claude, Cowork, and ChatGPT

These surfaces have no MemoryRouter lifecycle hook that forces a call on every turn. A skill or instruction can guide behavior, but does not convert it into automatic capture. Treat a successful tool result as the boundary; do not claim an unobserved write succeeded.

Plans and admin controls

Claude: Free supports one custom connector. Pro and Max support custom connectors. Team and Enterprise may require an Owner or Primary Owner to add/allow the connector and govern tools.
Cowork: requires a paid plan with Cowork access, subject to rollout and organization policy.
ChatGPT: Developer mode eligibility includes Plus, Pro, Business, Enterprise, and Edu in current general guidance, while plan-specific write capability still varies. Capability-detect. Business creation/publishing is admin/owner-controlled; Enterprise/Edu may delegate development with RBAC while admins/owners retain publication control.
Local coding/agent hosts: trusted-project, package, network, hook, and OAuth policy can be controlled by the machine or organization.

Vault and deletion boundaries

One OAuth connection binds to one vault selected during authorization. Conversational project handles are organization labels, not a server-enforced authorization boundary. Use separate vaults and connections for strict isolation.

forget_all_memories is whole-vault only. It needs the independent memories:delete scope and the exact phrase DELETE ALL MEMORIES. Per-memory deletion is available via the delete_memories MCP tool (same scope, explicit confirmation) or the dashboard.