Where agents remember your companyremember your company_

Memtro is a company memory behind an OpenAI-compatible API. You point the tools you already use at it with a per-user key; it injects relevant memories into each request, gives the model tools over your connected systems that read by default and act where you grant it, extracts durable facts afterwards, and can run recurring work on a schedule.

Free for 7 days, then one monthly price per organisation set by how many people use it: $50 for up to 10 users, $150 to 25, $350 to 50, $1,000 to 100, $2,500 to 250 and $5,000 to 500; above that we agree a custom plan. Connectors, jobs, memory and API keys are included at every size. A card is taken at sign-up and nothing is charged until the trial ends. Model usage is billed by your own providers under your keys; Memtro adds no margin.

Any time, from Dashboard → Organisation → Billing or the Stripe billing portal. Cancel during the trial and you pay nothing; cancel later and access continues to the end of the paid month with no further charge.

Only when you ask it to, and only through a tool marked as an action and granted by name. Every connector reads; most can also act, and the tools that change something — send a message or an email, open or update an issue, reply on a ticket, create or change a record, raise or settle an invoice, refund a charge, pause an ad campaign, publish a post, change somebody's role, buy a shipping label, create an event or a document — are marked on the Connections page, granted per team by an administrator (by its owner on a personal connection), and used by the model only when you ask for exactly that. Until a tool is granted, that connector only reads. Database queries run in a read-only transaction with a statement timeout and a row cap. Beyond that, Memtro writes only to its own memory.

In your organisation's memory, on servers Memtro operates in the European Union, with credentials encrypted at rest and backups taken by our hosting provider. Your requests go to the model providers you configured, under your own keys. See the Security page.

Anthropic, OpenAI, Google, xAI, DeepSeek, Meta and OpenRouter through your own keys, any OpenAI-compatible endpoint (Groq, Together, Ollama, vLLM), and two subscription engines — the Claude Code CLI with a Claude Code token, and Codex with a ChatGPT sign-in. Model auto routes by difficulty across a ladder you control, with a second ladder for software tasks.

Markers first and free — length, whether the message names anything of yours, whether tools are offered, whether it reads like reasoning or like a pleasantry. A decision provider answers it in a millisecond call if you have one. Only where neither is sure does a classifier run, on the cheapest rung of the ladder it is choosing within, so the decision can never cost more than the answer. Agent runs are held at the middle rung or above. Every reply carries the level, the ladder, the reason and the model that answered.

No. Every request runs on a key you hold, so the provider bill stays yours at your own rates, and the subscription is the only thing Memtro charges for. Routing exists to make that bill smaller: the cheapest rung that can answer each question, and a memory that answers from a handful of retrieved facts rather than from the thread and the documents people paste in today.

A provider can hold several keys — two subscriptions, an API key and a token, a separate account for one team — ordered as you like or set to spread requests across them at random. Routing tries every key of a rung before it drops to the level's second and third model, because a key that is rate-limited or out of credit is a problem with the key and not with the model. Handover only happens before any of the answer has been streamed; nothing switches mid-sentence.

Yes. Postgres, MySQL and SQL Server connections take an optional SSH tunnel: give Memtro a bastion host and a key or password and it forwards each query through it, so the database never needs a public address.

Yes. Create an API key that acts as its own agent: it becomes a member with its own personal memory, connections and jobs, gets the organisation default permissions plus any teams you add it to, and counts as one seat. Keys that act as you share your memory and take no extra seat.

Keys, connections, memory and jobs are either yours or the organisation's. Within a shared conversation the extractor classifies each fact, so knowledge about the company is shared while your own preferences and private matters stay personal. Personal conversations never write anything shared.

A normal request is one model turn with up to eight tool steps. Agent mode plans, gathers evidence across every relevant system, chases leads, verifies, saves what it learned, and answers once, ending with a line per system consulted. It is the Chat page's default.

Yes, both, per organisation and at no extra cost. Sign-in through OpenID Connect (Microsoft Entra ID, Okta, or any provider with a discovery document) or SAML 2.0, with your email domains verified by a DNS record, and an option that makes your provider the only way in — which is what makes removing somebody in your directory actually remove their access here. SCIM 2.0 handles joiners, leavers, renames and groups: groups become teams, marking somebody inactive withdraws their access while leaving their membership and memory intact, and a hard delete does whatever you chose in advance about what becomes of what that person alone knew. The provisioning token takes no seat and can read nothing.

Yes. Paste the token from claude setup-token under the Claude Code provider and use claude-code/opus, sonnet, haiku or fable. Requests run through a headless, sandboxed Claude Code CLI with Memtro's tools attached over MCP. Anthropic's consumer terms restrict subscription tokens outside Claude Code itself; it is your subscription and your call.

Yes, tool by tool. Each shared connection has a permissions page listing its tools, with a column for the organisation and one per team: tick what each may do. Slack, for example, separates searching messages, reading a channel and sending a message, so a team can read without ever posting. Nothing is usable until a tick exists, reading never implies acting, somebody in two teams gets everything either allows, and the rule holds in chat, the API, jobs, MCP and the sandboxed engines. The memories a connection records follow the same grants.

A prompt the agent runs on a cron schedule as its owner, with their tools and connections. Create one by asking in chat; every run's report is saved to a chat named after the job. Jobs can be personal or shared with the organisation.

Yes. Dashboard → Organisation → Usage shows requests and tokens for the last 7, 30 or 90 days or the current month, by user, by team, by model and by day, with a CSV export. Admins see the whole organisation; a team can be granted the same view; everyone sees their own.

If you sign in with a password, yes: every password account enrols an authenticator app (Google Authenticator, Authy, 1Password and the like) right after sign-up, and sign-in asks for a six-digit code after the password. Backup codes cover a lost phone and you can trust a device for 30 days. Google sign-ins are covered by Google's own second factor.