Memory infrastructure for your company's AI
Where agents remember your company
One API in front of the tools your team already uses and the models you already pay for. Memtro retrieves what matters, looks things up for real, learns from every answer, and runs the recurring work.
Sam was charged twice for order 4411: two payment captures a minute apart, one shipment, and Xero shows a duplicate paid invoice (INV-2202). Refund INV-2202 and reply on the ticket. BigCo is a £40k account owned by Priya, so flag it to her. help scout #2450 hubspot warehouse.payments xero INV-2202
Connects to the systems your business runs on
Every connector is read-only and purpose-built: the model gets the right tools for each system, and results are anchored into memory as people, companies, tickets and deals.
- Gmail
- Google Calendar
- Google Drive
- Postgres
- GitHub
- HubSpot
- Help Scout
- Xero
- Twenty CRM
- Slack
- Notion
- Linear
- Jira
- Confluence
- Stripe
Put your company's memory to work across every team
Answer tickets with the whole company behind you
Asked about a conversation, Memtro reads it, finds the customer in the CRM, checks their orders and invoices, and reports what it found and where. The agent cross-references by default; nobody has to ask twice.
Have a look at Help Scout ticket #2450
Walk into every call briefed
Account history, open tickets, invoices, the last email thread and what was decided internally, pulled together from the systems that hold them and remembered for next time.
Brief me on BigCo before the 3pm renewal call
Ask the ledger questions in plain English
Xero, Stripe and your warehouse answer through read-only tools. Duplicate charges, aged receivables, revenue by segment: one question, with the figures cited to their source.
Which invoices over £5k are more than 30 days overdue?
Keep decisions from evaporating
Decisions, owners and deadlines are extracted from conversations and tool results into a typed, time-aware memory. Corrections supersede old facts instead of piling up next to them.
Who owns the billing migration, and what did we decide about retries?
Better answers start with shared memory
The same question to the same model, with and without Memtro in front of it. Memory and live tools turn a plausible paragraph into a cited answer.
What do I need to know for my 3pm with BigCo?BigCo renewed for £40k on 2 October (owner Priya Sharma). There is an open Help Scout ticket, #2450, about a duplicate charge on order 4411: the warehouse shows two payment captures and Xero a duplicate paid invoice, INV-2202, which has not been refunded yet. Sam Lee, your contact, emailed on Tuesday asking for a timeline. Lead with the refund.
What do I need to know for my 3pm with BigCo?I don't have access to your CRM or ticketing system, but here is a general checklist for a customer meeting: review the account history, check for open support issues, confirm the renewal date, and prepare questions about their goals for the coming quarter…
Designed to compose
Ask, investigate, remember, schedule: one conversation, one memory, the same tools everywhere.
- ask:
Who owns the billing migration?
- investigate:
Have a look at Help Scout ticket #2450
- remember:
Remember that Acme renewed for £40k
- schedule:
Every weekday at 9, summarise VIP tickets older than a day
Production-grade trust
Every memory carries where it came from, every connector is read-only, and every agent engine runs in a sandbox. The organisation's knowledge and each person's own notes are kept apart by design, not by policy.
- Provenance on every factmemories carry the tool, system, record id and URL they came from
- Personal versus shared, classifiedyour preferences never leak into the organisation's memory
- Read-only connectorsqueries run in read-only transactions with timeouts and row caps
- Credentials encrypted at restprovider keys and connection tokens use AES-256-GCM
- Sandboxed enginesOpenCode and Claude Code run in one-shot containers with no access to the host
- Your keys, your databasebring your own provider keys; self-host on one Postgres
Everything your agents need to work with your company
From a single question to always-on jobs, Memtro gives agents every way to find, remember and act on what your organisation knows. All on one platform.
Chat API
→The OpenAI chat-completions shape at /v1, so Cursor, Open WebUI, the SDKs and curl work unchanged. Per-user keys, streaming, tools.
Memory
→A knowledge graph and vector index in Postgres. Facts, decisions, events and preferences with provenance, personal or shared, superseded when they change.
Connectors
→Read-only tools for Gmail, Postgres, GitHub, HubSpot, Help Scout, Xero, Slack, Notion, Linear, Jira, Confluence, Stripe and more, with hourly sync.
Routing
→Model auto picks a rung on a ladder you control by how hard the question is. Agent mode never routes below the middle rung.
Agent mode
→Plan, query every relevant system in parallel, cross-reference, verify, save what was learned, answer once with a line per system consulted.
Jobs
→Scheduled actions created from chat or the dashboard. Each run has the owner's tools and lands its report in a saved chat.
MCP server
→Every memory and connector tool available to Claude Code, OpenCode, Cursor and Claude Desktop with one URL and a key.
Engines
→Run agent mode on the native loop, on OpenCode, or on Claude Code, including your own Claude Code subscription, sandboxed in Docker.
More signal per token
Model auto scores each request and picks a rung on a ladder you control: the cheapest model for a quick lookup, the strongest for a real investigation. Agent runs never land below the middle rung. Every response says which model answered and why.
| level | example | anthropic ladder |
|---|---|---|
| 1 | thanks! | haiku 4.5 |
| 2 | what is Sam's email? | haiku 4.5 |
| 3 | look at ticket #2450 (agent) | opus 5 |
| 4 | design the retry policy | opus 5 |
| 5 | model next year's cash flow | fable 5.1 |
Pay for models, not for a middleman
Memtro uses the provider keys you already have, so model usage is billed by your providers at their rates, or by your Claude Code subscription. Everything Memtro stores lives in one Postgres database you can export or host yourself.
- Per-user API keys, personal and shared resources
- Unlimited connectors, jobs and saved chats
- Self-host on a single server with the open deployment scripts
Start for free
Sign in, add a provider key, create an API key, point your tools at Memtro. About five minutes.
Working in Claude Code, OpenCode, Codex or Cursor? Paste this and your agent connects itself.
curl -s https://www.memtro.com/agents.md
Available everywhere you build
FAQs
What does Memtro do?+
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 read-only tools over your connected systems, extracts durable facts afterwards, and can run recurring work on a schedule.
Which models can I use?+
Anthropic, OpenAI and Google through your own keys, any OpenAI-compatible endpoint (Groq, Together, OpenRouter, Ollama, vLLM), and the Claude Code CLI with a Claude Code token. Model auto routes by difficulty across a ladder you control.
Does Memtro change my data?+
No. Every connector is read-only. Postgres queries run in a read-only transaction with a statement timeout and a row cap. Memtro writes only to its own memory.
What is the difference between personal and shared?+
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.
How does agent mode differ from a normal request?+
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.
Can I use my Claude Code subscription?+
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.
What is a job?+
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.
Can I self-host?+
Yes. One Ubuntu server with Postgres, nginx and Docker runs everything; there is no external database or auth service. The self-hosting guide covers setup, environment, timers and backups.