Vertiso Memory
Give every AI tool one durable memory store: facts, decisions, preferences and session handoffs that survive across Claude, Codex and Cursor. Anthropic lists 14 tools; Vertiso's own MCP reference page is a placeholder, so two of the 14 are documented nowhere. OAuth with a single wildcard scope.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Connect Vertiso Memory via MCP
https://memory.vertiso.ai/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Vertiso Memory Tools & Capabilities (14)
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
Limits
- There is no vendor tool reference. Vertiso's documentation nav links MCP Reference, CLI Reference and API to one placeholder page. The 14 names come from Anthropic's directory, and two of them —
forgetandunarchive_memory— appear in no Vertiso document we could find. This is the "no vendor enumeration" case, which is weaker than a clean diff: there is no second list to check the first against. - The tool list is a floor. Vertiso's own shipped skills call
complete_action,list_actionsandwork_status, none of which Anthropic lists. We did not verify them against the live server, because doing so would require calling a tool. complete_actiontoggles when called without an argument. Vertiso's wrap-up skill warns that passing neitherclosednorstatusflips the current state and "would reopen an action already closed elsewhere." That warning exists only inside a skill file; an MCP-only client never sees it.- The only OAuth scope is a wildcard.
mcpnames no object and no verb, so consent cannot separate recalling a memory from forgetting one. The boundary a reader would most want — between the routine write and the irreversible one — does not exist. - Secret stripping is an instruction to a model, not a filter. Vertiso says so itself: automated detection "can miss sensitive material, so do not rely on it as a substitute for sharing carefully." If you point a capture skill at a repository whose tool output carries credentials, redaction depends on the model noticing.
- The CLI installer writes a hook the marketplace plugin does not. Vertiso's guarantee about no automatic transcript upload is scoped to marketplace plugins.
vmem mcp installcan add a user-scoped SessionStart hook, extra skill copies and a second MCP entry. Do not run it for a client that already has the plugin. vmeminstalls by piping a remote script to a shell.curl -fsSL https://memory.vertiso.ai/install.sh | shis Vertiso's documented path.- Free-tier caps are small. 100 stored memories at 4,000 characters each. No tool reads the remaining count, so an agent cannot check before it writes.
- Revision history, shared team memory, SSO and audit logs are not available. They sit on Teams, marked "Coming soon", and Enterprise, marked "Contact us".
- Memory content reaches Google Gemini. Vertiso names it as the subprocessor for memory structuring, retrieval and embeddings, stating Google does not train on paid-API submissions. Enterprise bring-your-own-inference is listed as not included on every self-serve plan.
- We could not read tool schemas or safety annotations. The endpoint returned 401 to an anonymous request on 2026-08-23. Every parameter named on this page comes from Vertiso's skill sources, not from the server.
- Retention is indefinite by default and deletion is not self-service for the account. Individual memories can be deleted from the web app, but account and Memory Data deletion is an email request to Vertiso, completed within 30 days and unrecoverable once begun.
- The
robots.txtcarries aContent-Signalofsearch=yes, ai-train=no, use=reference. It expresses noai-inputpreference either way. We honouredai-train=no, treateduse=referenceas the reason to cite and link rather than reproduce, and synthesised no vendor prose verbatim.
Frequently asked questions
Durable facts your assistant decides are worth keeping, plus anything you explicitly ask it to save. Vertiso's guide names the memory types: decision, project, constraint, action, intent, observation, event and lesson. Its skills also send attributed verbatim excerpts of your conversation when exact wording preserves a decision better than a paraphrase would.
One of the 14 tools is named forget, and Vertiso documents no such tool anywhere. Its guide distinguishes archiving, which keeps a memory but marks it inactive, from deleting, which removes it for good, and the vmem CLI ships an explicit vmem rm command. We could not read the tool schema, so treat forget as irreversible until Vertiso publishes its reference.
Exactly one, and it is a wildcard. The server's RFC 9728 descriptor, its authorization server metadata and its 401 challenge all advertise a single scope named mcp on 2026-08-23. It names no object and no verb, so granting it grants whatever the server implements now and whatever it implements later.
No. The only hook the Claude plugin ships is a SessionStart shell script that runs vmem hello with a source and a client name, then exits zero. We read its source on 2026-08-23. Vertiso's Trust page states the marketplace plugin installs no pre-compaction transcript hook and does not continuously upload conversations.
Yes, when you invoke one. The checkpoint skill instructs the agent to read the Claude Code .jsonl session transcript and quote from it rather than from working memory, so quotes stay exact. All three capture skills require you to invoke them or accept an offer first, and each carries a secret-stripping step covering tokens, keys and connection strings.
Anthropic lists 14 and Vertiso publishes no list of its own. Its documentation nav links MCP Reference to a placeholder page reading Coming soon. Vertiso's own plugin skills additionally name complete_action, list_actions and work_status, which are absent from the 14, so 14 is a floor rather than a total.
No, and its privacy policy commits to that twice. It states Vertiso does not use your Content or Memory Data to train, fine-tune or otherwise develop models. Memories do reach Google Gemini for structuring, retrieval and embeddings, under paid-API terms Vertiso states exclude training. Google is one of seven named subprocessors.
Indefinitely while your account exists. Vertiso's retention table gives Content and Memory Data no expiry during an active account, and keeps it on the Free plan after a paid subscription ends. Deletion happens within 30 days of a confirmed account deletion, and Vertiso states data is inaccessible and unrecoverable once deletion begins.
Sources
- Vertiso Memory install and connection guide, all client paths, the vmem CLI and the bootstrap instruction (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory usage guide, memory types and the recall-versus-search split (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory Trust centre, capture model and control claims (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory documentation placeholder, reached from the MCP Reference, CLI Reference and API nav links (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory machine-readable index,
llms.txt(retrieved 2026-08-23).llms-full.txtreturned 404 and the.mdsuffix route returned 406. · retrieved 2026-08-23 - Vertiso Memory pricing and plan comparison (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Corporation privacy policy, subprocessor table and retention table (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory Acceptable Use Policy, a negative enumeration of prohibited uses (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso Memory security and support pages — · (retrieved 2026-08-23) · retrieved 2026-08-23
- Claude plugin repository, MIT, the SessionStart hook source and its manifest — · hook at
vertiso-memory/hooks/session_start.sh, wiring atvertiso-memory/hooks/hooks.json(retrieved 2026-08-23; last pushed 2026-08-22) · retrieved 2026-08-23 - Bundled skill sources —
checkpoint,handoff,handoff-resumeandwrap-upundervertiso-memory/skills/, byte-identical in , and (retrieved 2026-08-23). Thecomplete_action,list_actionsandwork_statusnames are in the wrap-up skill. · retrieved 2026-08-23 - Codex plugin manifest with the vendor's default example prompts —
vertiso-memory/.codex-plugin/plugin.jsonin (retrieved 2026-08-23) · retrieved 2026-08-23 - JetBrains plugin, skills only and no hooks (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso engineering blog, "What it costs to remember", token and accuracy comparison (retrieved 2026-08-23) · retrieved 2026-08-23
- Vertiso engineering blog, "Ninety-three point four, honestly", the 93.4% LongMemEval run (retrieved 2026-08-23) · retrieved 2026-08-23
- Live OAuth posture check: 401 with a
WWW-AuthenticateBearer challenge carryingscope="mcp"at the endpoint, RFC 9728 metadata athttps://memory.vertiso.ai/.well-known/oauth-protected-resourceand/mcp, and authorization server metadata athttps://memory.vertiso.ai/.well-known/oauth-authorization-server(2026-08-23). A control path under the same well-known prefix returned the site's ordinary 404, so the descriptor is genuine rather than an interpolating catch-all. · retrieved 2026-08-23 - Anthropic Connectors Directory entry — , read from our directory snapshot dated 2026-08-23 · retrieved 2026-08-23
- Vertiso Memory support — · Privacy
Use in Agentman
Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.
Open in Agentman StudioServer Info
- Category
- Productivity
- Developer
- Vertiso Memory
- Tools
- 14
- Domain
- memory.vertiso.ai
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.