AI for financial services, with every figure traced to its record
A payments, lending or fintech business has the data; the problem is that a question about a customer touches the payment processor, the bank, the ledger and the warehouse, and the answer must be right. Memtro connects to those systems through connectors that read by default, runs warehouse queries inside read-only transactions with the SQL shown, and cites every figure to the row, charge or invoice it came from.
7-day free trial · card on file, nothing charged until it ends · cancel any time
Clients, accounts, transactions, documents and the ledger answered from one memory, with every figure cited to its record.
The systems these questions draw on, connected to one memory. Every connector reads by default; the tools that change something are granted by name, per team.
What it looks like today
- 01A reconciliation question bounces between finance, operations and engineering for a day.
- 02Reports are rebuilt by hand for every ad-hoc question from a regulator or a board.
- 03Nobody wants to give an AI write access to anything that moves money.
- 04The same figure differs between the ledger and the warehouse, and nobody knows which is right.
- 05A board question arrives on Thursday for a Friday meeting, and the answer needs three teams to assemble it.
- 06Every new analyst learns which table is the right one by asking the person who built it.
What Memtro does about it
One memory in front of the systems you already pay for, reading by default and acting only where an administrator has granted the tool by name.
- Connect Stripe, GoCardless, PayPal, Wise, Starling, Brex or Ramp, NetSuite and the warehouse on Snowflake or Postgres, reading by default; nothing in this set acts until an administrator grants a tool by name.
- Warehouse questions run as bounded SQL inside read-only transactions with a row cap and a timeout, and the SQL is part of the answer.
- Contradictions between two systems are found overnight and put to a person, whose answer corrects memory itself.
- Permissions are granted tool by tool to named teams, so the people who may see card spend are the people who should.
- Schedule jobs: failed and disputed payments every 15 minutes, unreconciled transactions daily, a monthly summary of fees by provider.
System by system
What Memtro reads in each part of the stack, and what it may change there once the tool has been granted.
Payments
Stripe, GoCardless, PayPal, Square and Wise: charges, mandates, refunds, disputes, payouts and fees, read only, matched to the customer they belong to.
Banking and spend
Starling Bank, Brex and Ramp: transactions, balances, card spend and who spent it, beside the ledger entry that was supposed to match.
The ledger
NetSuite, Xero, QuickBooks Online, Sage Accounting or Dynamics 365 Business Central: invoices, bills, journals and the month's position.
The warehouse
Snowflake, BigQuery, Redshift, ClickHouse, Athena, Postgres, MySQL, SQL Server or MongoDB: schema first, then a bounded read-only query with a row cap and a timeout, and the SQL shown in the answer.
Operations and monitoring
Datadog, Grafana, Sentry, PagerDuty, incident.io, CloudWatch and CloudTrail, so a spike can be read as a customer problem or a system one rather than guessed at.
The public record
Companies House: officers, filings, accounts and registered charges for a counterparty or a corporate customer, answered inside the conversation.
There are 144 connectors in all, each with its own tools, its own permissions and a page of its own. Browse every connector →
Who asks what
Finance lead
Asks the ledger and the processors the same question at once, and gets an answer that says where each figure came from and where the two disagree.
Payments operations
Sees failures, disputes and stuck settlements as they happen rather than at month end, with the customer and the last contact attached.
Analyst
Gets the schema explained and the query written, then keeps both: memory holds which table answers which question, so the next analyst starts where this one finished.
Risk and compliance
Asks what was known and when. Memory can be read as it stood on any past date, every fact names its record, and usage by person and team is exportable.
The month, end to end
- 01
Every quarter of an hour
Failed and disputed payments in the last hour arrive with the customer and the amount, so the first person to know is not the customer.
- 02
Every morning
Ledger entries with no matching settlement older than two days, and anything the overnight contradiction check found between two systems, put to a person whose answer corrects memory itself.
- 03
Month end
Fees by provider, volumes by customer and the ledger against the processors, each figure cited and each query shown so the next month is a rerun rather than a rebuild.
- 04
On request
A board or regulator question is answered from the warehouse with the SQL in the answer, which means it can be checked, rerun and handed to somebody else unchanged.
Ask it like this
Plain English, from chat, your editor, a scheduled job or your own agent. In agent mode each question also checks the other systems that know about the same people, companies and references, and the answer ends with a line per system consulted.
- Stripe
- GoCardless
- PayPal
- Wise
- Starling Bank
Every figure in the answer cites the record it came from, and the run ends with a line per system consulted.
- ›Revenue by product line for Q3 from the warehouse, with the SQL you used
- ›Which GoCardless mandates failed in the last 24 hours, and have we told the customers?
- ›Does NetSuite agree with Stripe on last month's card revenue?
- ›Which customers had a payment fail twice this month, and what did we do each time?
- ›Show card spend by team last month from Brex, with the five largest items
- ›Which Companies House filings changed for our twenty largest corporate customers this quarter?
- ›What did we know about the settlement discrepancy as of the first week of September?
Work that runs itself
A job is a prompt the agent runs on a cron schedule as its owner, with their tools and connections. Say any of these in chat and the job exists; every run's report is kept in a chat you can question.
- schedule:Every 15 minutes: failed or disputed payments in the last hour, with the customer and amount
- schedule:Every day 07:00: transactions in the ledger with no matching settlement, older than two days
- schedule:First of the month: fees paid by provider and the three largest customers by volume
- schedule:Every Monday 07:00: balances and movements across every bank and payment account, with anything over the threshold you set
- schedule:Every quarter: customers whose payment volume halved, with their last contact and any open dispute
Questions people ask
Can it move money?+
No. A connector acts only through tools marked as actions, granted by name per team and used when a person asks, so nothing moves money unless you have granted exactly that tool. Database queries run in read-only transactions with timeouts and row caps, and the agent can only call tools, never a shell.
Where does the data live?+
In your organisation's memory, encrypted at rest, with provider keys and connection tokens under AES-256-GCM. You bring your own model keys, so prompts go to the provider you already have an agreement with.
Can we prove what was deleted?+
Yes. A right-to-erasure request removes everything held about a person across every store, verified in the same transaction and receipted.
Can we see the SQL it runs?+
Always. The query is part of the answer, so it can be read, rerun and corrected. Queries run inside a read-only transaction with a statement timeout and a row cap, under the credentials you gave — usually a replica with a restricted role.
Can it reach a database on a private network?+
Yes. Postgres, MySQL and SQL Server connections take an optional SSH tunnel: give Memtro a bastion host and a key or password and every query is forwarded through it, so the database never needs a public address.
Which models see our data?+
Only the providers whose keys you hold, under your own agreements and at your own rates, with no margin added. The ladder of models the router may use is yours to set, per team, and every answer says which model answered and why.
The same work, by team
The same connectors and the same memory, described through the job somebody does rather than the business they do it in.
- FinanceAsk the ledger questions in plain EnglishXero, Stripe and the warehouse answer through tools that read by default, and Xero can raise a draft invoice or record a payment when you ask and an administrator has granted it. Duplicate charges, aged receivables, revenue by segment: one question, with the figures cited to their source.Memtro for finance
- Data and analyticsQuestions answered from the warehouse, with the SQL shownPoint Memtro at a read-only replica and anyone can ask the warehouse a question in plain English. The agent reads the schema, writes the query, runs it inside a read-only transaction, and shows its work.Memtro for data and analytics
- LeadershipOne question across the whole companyFounders and managers ask what is happening and get an answer stitched from support, sales, finance and engineering, with every figure traced to its source and the rest of the week handled by jobs.Memtro for leadership
Ask it about your own business
Connect the first two systems, ask the question you opened this page with, and see where the answer comes from. About five minutes.
7-day free trial · card on file, nothing charged until it ends · cancel any time