Tables: give Claude a database, not just a memory
Claude's memory carries the context of your past chats. It was never built to keep records. Tables are structured records your AI creates, fills from the tools you already use and reads back in any chat after, with typed fields, a key that stops duplicates, and every row editable in your dashboard.
On Monday you ask Claude for every invoice more than 30 days overdue. It reads Xero, finds nine, and lays them out with amounts, contacts and who to chase first. It’s a good list.
On Tuesday, in a new chat, you ask for the list again. Claude remembers you were chasing invoices. What it gives you is a new list, built again from Xero, with the three you chased yesterday back at the top, because nothing wrote down that you chased them.
A chat is a scratchpad. The moment you want a list that lasts, there has been nowhere for it to go.
Tables is that place. They’re structured records your AI client creates, writes and reads through FloConnector. They stay put between chats, your team sees the same rows, and you can open and edit every one of them in the dashboard. The one-minute version:
Claude memory is for context, not records
Claude has memory, and it’s good at what it’s for. It carries context from one chat to the next, and on paid plans it can search your past conversations when you ask. It knows who you are, what you’re working on and how you like things written, so you stop explaining yourself every morning.
What it isn’t is a place to keep records. Memory holds what it judged worth keeping, in its own words. That is the right design for context and the wrong one for a list of nine invoices with amounts on them:
- It recalls, it doesn’t read. Ask for Tuesday’s exact list and it’s working from memory. Usually close. Never something you’d send to your accountant.
- It can’t be sorted, filtered or totalled. There are no columns to sort on and no amounts to add up.
- It’s stuck in one person’s chat. The list your office manager built is not something your bookkeeper can open.
ChatGPT’s memory works on the same idea, with the same limit. Neither is a bug. Context and records are different jobs, and the second one needs a database.
| Claude's memory | A spreadsheet connector | A table | |
|---|---|---|---|
| Built for | Context across chats | A grid people edit | Records an AI writes |
| The exact list, next week | Recalled | Read | Read |
| Text in a number field | Not its job | Accepted | Refused, with the reason |
| The same lead added twice | Not its job | Two rows | One record, matched on the key |
| A batch that fails halfway | Not its job | Half written | Nothing written |
A spreadsheet is a fine answer when your team already lives in one, and the Google Sheets and Airtable connectors are there for exactly that. The trouble starts when the AI becomes the main writer. A model writing to a sheet has to keep the grid in shape itself, and nothing in the sheet says no.
What a table is
A table is a list of fields and the records that fill them. Every field has a type, and there are eleven: text, long text, number, checkbox, date, date and time, a single option, several options, email, URL and phone. A value that doesn’t fit is refused with the reason, never guessed at. 1,5 is not a number. A time without a timezone is not a time.
You don’t design any of this. You say what you want to keep track of, and your AI makes the table, picks the fields and fills it:
YouStart a chase list: every Xero invoice more than 30 days overdue, with who we've emailed and when.
ClaudeDone. I've made a table called Invoice chases, keyed on the invoice number, with contact, amount due, due date, last chased and status. 9 invoices, $18,420 outstanding. None have a chase date yet.
YouA week later, in a new chat. Who haven't we chased in the last seven days?
ClaudeThree: Harbour Freight (record 4, $2,310), Northwind (record 7, $4,880) and Boltworks (record 9, $1,150). $8,340 between them.
The second chat didn’t rebuild anything. It read the table.
Three things make that work, and they’re the difference between a table and a note:
The key. One field can be named the key, the thing that makes a record unique: an email, an invoice number. Two records can never share a key value, and matching ignores case, so Ann@Acme.io and ann@acme.io are the same person. Ask your AI to “add these leads” twice and you still have one record per lead.
Record numbers that never move. Records are numbered 1, 2, 3 in the order they arrive, and a number is never reused. Delete record 12 and the next one is still 13. “Record 4” in a note next month means the same record, or nothing, but never a different one.
Rows you can see. Every table sits in your workspace beside your connectors and skills. Open one and it’s a grid: sort, filter and search on any field, fix a cell, add a field, export the lot to CSV whenever you like.
Built for a model to write to
Models retry when a call is slow. They repeat themselves, and they guess. A database for AI has to assume all three, so every rule below exists to keep them out of your data.
All or nothing. Your AI writes up to 500 records at a time, and every one is checked before any is saved. If one is wrong, nothing lands, and the error names each bad record and why. The AI fixes those and sends the batch again, and because nothing landed the first time, nothing is doubled.
No silent overwrites. Every record carries a version that goes up each time it changes. Your AI can ask for a write to be refused if someone changed the record since it last looked, so two chats can’t quietly overwrite each other.
Changes never rewrite your records. Turn a text field into a number and a record holding “12 units” keeps it. The change reports how many records no longer fit, and they stay as they are until someone edits them.
The destructive steps need a person. Removing a field deletes its values from every record, so your AI has to ask you before it does it. Deleting a whole table only happens in the dashboard, behind a confirmation. Your AI can archive a table, which freezes it, and restore it again.
The ten tools
Tables arrive as ten tools on the FloConnector connector, FloConnector’s own tools, alongside the ones for writing skills.
| Tool | What it does |
|---|---|
flc_table_list | Every table, its record count and fields, and how much room the plan has left |
flc_table_describe | One table's fields, types, key and two example records |
flc_table_create | A new table with typed fields and an optional key |
flc_table_update | Rename, add, change, remove or reorder fields |
flc_table_archive | Archive a table, or restore it |
flc_record_query | Filter, sort, search, count and page through records |
flc_record_insert | Add up to 500 records at once |
flc_record_upsert | Add or update up to 500, matched on the key |
flc_record_update | Change fields on records by number |
flc_record_delete | Delete records by number |
| 10 tools | $0 per call, on every plan |
They cost nothing per call because they work on FloConnector’s own data rather than a vendor’s API. You only pay for calls that reach a connected platform, like the Xero read that filled the chase list.
Fill it from the tools you already use
A table on its own is a list you could have typed. It gets useful when your AI fills it from everything else it’s connected to.
Point it at your inbox, your booking page and your forms. It reads each one through its connector, matches what it finds on the key, and writes the lot to one table. Next time, it reads the table instead of the inbox. Ask on a schedule, which Claude and ChatGPT both support now, and the table keeps itself current: every morning, new quote requests from Gmail and Typeform, matched against who’s already on the list.
Some of the jobs people use tables for:
- A lead list built from enquiry forms, booked calls and email, one record per person
- A content calendar with a status for each post, so “what’s going out this week” has one answer
- An invoice chase, like the one above, with who was emailed and when
- A job log or a decision log, so a choice made in March is still findable in June
For anything bigger than a quick answer, a total by month or a table joined against invoices from your accounting connector, your AI can load the whole table into a temporary, in-memory analytics workspace and query it with SQL. The rows don’t pass through the conversation, only the answer does. How that works, and why it isn’t a code sandbox.
Who can see a table
Owners and admins of the workspace. Members see none by default, not even to read, because a table can hold customer details and amounts, and a new teammate joins as a member. A custom role can give members read or write access.
When your AI uses a table it acts as you, with your permissions. Someone who can read tables but not change them gets exactly that through their AI, and nothing more.
Limits that never delete
Every plan includes tables. The limits are per workspace, with records counted across all of its tables:
| Plan | Tables | Records |
|---|---|---|
| Solo | 1 | 1,000 |
| Pro | 25 | 50,000 |
| Scale | 100 | 250,000 |
A write that would go over a limit is refused, and the message says how many records would still fit. If a workspace ends up over a limit, after moving to a smaller plan for example, nothing is deleted, ever. Its tables become read-only until it’s back under: you can still view, export and delete, but nothing can be added or changed. See pricing for the plans.
Connectors, skills and tables
Connectors let your AI act in your tools. Skills tell it how you want the job done. Tables give it somewhere to keep what it finds.
They live in one place, behind one endpoint per profile, and they work with any AI client that supports connectors: Claude, ChatGPT, or whatever comes next. If “connector” or “MCP” is new to you, there’s a plain-English explainer.
How to start a table in Claude
- Create a FloConnector workspace
- Open Connectors, find FloConnector and choose Enable. There’s no login: it’s FloConnector itself
- Add it to the profile your AI client uses, the same way you add any connection
- If that profile isn’t in Claude yet, copy its endpoint into Claude as a custom connector. If it is, disconnect and reconnect so Claude picks up the new tools
- Say what you want to keep. “Start a table of everyone who asks for a quote, with their company, budget and where we’re up to” is a good first one
You can also make a table yourself in the dashboard under Tables: give it a name, add columns, and pick a key. The Tables help page covers every field type and rule.
Common questions
Claude already has memory. Why would I need this? Memory keeps context: who you are, what you’re working on, how you like things done. A table keeps records: the exact rows, with types, that you and your team can sort, total and edit. Most people want both, and they don’t overlap.
Is this an MCP database server? Not in the sense of connecting your own Postgres or MySQL to Claude. It’s a database FloConnector hosts for you, reached through ten MCP tools, with nothing to install, run or back up.
How is it different from the Google Sheets connector? A sheet is a grid a model has to keep in shape itself. A table has typed fields, a key and all or nothing writes, so a model can’t put text in a number column, add the same lead twice, or leave a batch half written. If your team already works in a sheet, keep using it. When the AI is the main writer, use a table.
Does it work with ChatGPT? Yes. Tables are tools on a standard connector, so any client that supports custom connectors can use them, and a table one client wrote is readable from the other.
Can my AI delete a table by accident? No. It can delete records you ask it to delete, and it can archive a table, which can be undone. Deleting a table only happens in the dashboard, behind a confirmation.
Can I get my data out? Every table exports to CSV from the dashboard, and your AI can read every record through the connector at any time.
What does it cost? Nothing per call, on every plan. The plan sets how many tables and records a workspace can hold, see pricing.
Is it ready to use? Tables is an early release, marked Alpha in the dashboard. It works and it’s safe to use. How tables are structured may still change as we learn from real workspaces. The full tour is on the Tables page.
Your AI already finds things. Now it has somewhere to put them.
Keep reading
More from the blog
Pylon + AI: 25 tools, and the signed quote stops being retyped
Pylon runs on a connector we build and maintain ourselves, with 25 tools across projects, designs, quotes, hardware prices, materials and STCs. The quote a customer signs can now be read line by line, priced off your own catalogue, and checked against what the proposal actually said.
Xero MCP: 133 tools, and month-end is something you can ask for
Our Xero connector went from 53 tools to 133. It can now send the invoice, hand you the link your customer pays from, put an overpayment against the next invoice, record a twelve-invoice remittance as one payment, and read the reports your accountant opens. Here's what you can ask for now.
Connecting someone else's accounts to Claude, without their password
Every guide to MCP authentication assumes the person connecting the account is the person running the AI client. When you are doing the work for somebody else, they are not. This is the thing I could never solve properly in n8n, and what it took to solve it here.