Skip to content

What AI actually does with write access to your notes: 108,613 tool calls analyzed

Everyone argues about whether you should let an AI write to your notes. Almost nobody has data on what happens when you do.

We do. Hjarni is a knowledge base with a built-in MCP server, which means every tool call an assistant makes is a line in our metrics. Here is what the last 30 days actually look like: 108,613 MCP calls from 558 people who connected ChatGPT, Claude, and other assistants to their notes. About one call every 24 seconds, around the clock.

Quick answer

Given read and write access to a knowledge base, AI assistants spend 54% of their calls reading and searching notes and 24% writing. They update existing notes 3.7 times for every new note they create, fetch the user's custom instructions 9,480 times a month before acting, and check a note's edit history 2,320 times. Only 0.9% of calls error. The AI behaves less like a chatbot with a notepad and more like a careful co-author.

The shape of a month

The ten busiest tools, out of 108,613 total calls:

Tool What it does Calls
notes-get Read one note in full 30,922
search Search notes, folders, tags 25,587
notes-update Edit an existing note 20,816
instructions-get Fetch the user's AI instructions 9,480
notes-create Write a new note 5,694
me Look up account and context 2,970
notes-list Browse notes in a folder 2,527
notes-history Read a note's edit history 2,320
containers-create Create a folder 1,035
dashboard-get Fetch the overview 975

Reads (fetching, searching, browsing) are 54% of everything. No surprise: retrieval is the point of connecting a knowledge base. The writes are the surprise.

A quarter of everything the AI does is writing

26,510 calls in 30 days created or changed a note. That is 24% of all traffic. People do not connect their notes as a read-only reference. They let the assistant close the loop: research something, then save it. Decide something, then log it. Finish a meeting, then file it.

This is the part of the AI memory conversation that usage data settles. The debate is framed as "should AI be allowed to touch your notes." The people who actually connected one answered with 26,000 writes a month.

It maintains more than it hoards

The common fear is note sprawl: give an AI write access and it buries you in generated junk.

The data shows the opposite ratio. Assistants updated existing notes 20,816 times and created new ones 5,694 times. That is 3.7 updates for every create. Connected to a structured knowledge base, the AI mostly keeps existing notes current: appending a decision, correcting a stale fact, extending a log.

That behavior is not automatic. It comes from the assistant being able to search before it writes, and from instructions that say things like "update the existing note instead of creating a duplicate." Which brings us to the next number.

It reads the house rules first

The fourth most-used tool in the entire dataset is instructions-get: 9,480 fetches of the user's custom AI instructions in a month. Before assistants write, they pull the rules the user set for that folder or team: what format, what tags, what tone, what not to touch.

Instructions are the difference between an AI that writes into your system and an AI that writes over it. The usage says assistants take them seriously.

It checks its own history

Assistants read a note's edit history 2,320 times in 30 days: who wrote what, when, and what changed. An AI double-checking provenance before trusting or editing a note is exactly the behavior you want from something with write access.

And when a write goes wrong, it is visible and reversible. Every edit in Hjarni is versioned and attributed, human or AI, with one-click revert. Across the full month, 0.9% of calls returned an error, and the errors we read through are reassuringly boring: a mismatched edit anchor, a version conflict when two sessions touched the same note, a wrong ID. Each of those kinds is rejected before it touches the note, so the failed call leaves your notes exactly as they were.

ChatGPT is where MCP actually landed

One more number worth being honest about: ChatGPT made 85% of these calls and represents 92% of the people using MCP here, with Claude a distant second and coding tools like Claude Code a rounding error. One product's sample, but a strong signal about where MCP adoption actually landed after ChatGPT shipped connector support.

The caveats

This is one product, one month, aggregate data only. The window is the 30 days ending 2026-08-25, counted from Hjarni's own MCP server logs: every completed tool call from a connected client, no sampling, with our internal test accounts excluded. The sample is people who chose to connect an assistant to a knowledge base, which selects for intent. And event counts describe behavior, not quality: the ratio says the AI edits more than it creates, not that every edit was good. That is what the version history is for.

What to do with this

If you have been on the fence about write access, the data describes what the other side looks like: mostly reading, disciplined writing, rules checked first, history checked often, everything undoable.

Give your AI a memory and see your own numbers. Setup takes five minutes for ChatGPT or Claude. And if the safety question is the one holding you back, we wrote up exactly what a connected assistant can and cannot do.

Give your AI a memory. Free.

Connect Claude or ChatGPT to notes they can actually read and write.

Get started free

Give your AI a memory. Free.

Common questions

FAQ

What does an AI assistant actually do with access to a knowledge base?

In 30 days of Hjarni MCP traffic, 54% of calls were reads (fetching and searching notes), 24% were writes (creating and updating notes), and 9% were the assistant fetching the user's custom instructions before acting. The rest was housekeeping: listing folders, checking note history, and account lookups.

Will ChatGPT or Claude mess up my notes if I give them write access?

The data says it is rare. Across 108,613 calls, 0.9% returned an error, and the errors we read through were refused writes rather than damaged notes: a mismatched edit anchor, a version conflict between two sessions, a wrong note ID. Those are rejected before they touch the note. Assistants also checked note history 2,320 times in the same window. Every edit in Hjarni is versioned and attributed, so a bad edit is one click from undone.

Does AI clutter your knowledge base with junk notes?

Less than people expect. In our data, assistants updated existing notes 3.7 times for every note they created: 20,816 updates against 5,694 creates in 30 days. Connected to a knowledge base with structure and instructions, AI behaves like a maintainer, not a hoarder.

Do AI assistants follow custom instructions in a knowledge base?

They ask for them constantly. Fetching the user's AI instructions was the fourth most-used tool in our data, with 9,480 calls in 30 days. Assistants pull the house rules before they write, which is exactly what per-folder instructions are for.

Which AI assistant uses MCP the most?

In Hjarni's data, ChatGPT: 85% of all MCP calls and 92% of active MCP users in the last 30 days, with Claude a distant second. That is one product's sample, but it matches where MCP adoption went after ChatGPT added connector support and the Apps Directory.

Start here

Write once. You both remember.

Free to start. No credit card required.

5 out of 5 from 5 reviews 1,700+ notes a day read or updated by AI

Works with Claude and ChatGPT today.