My AI forgets between sessions. Notes fix that.
Every agent I work with is brilliant for one session and a blank slate the next. Close the window and the reasoning, the gotchas, the "we already tried that" all vanish. I run a dozen sites this way, so that amnesia used to cost me the same explanation over and over. Now every project keeps its own notes in askbowtie, and picking a project back up is a read, not a re-brief.
The short version
Notes are a shared, per-domain log your agent writes to and reads from over MCP. They are how a project remembers itself across sessions, agents, and months.
- Every domain writes its own. The vpn.com agent, the askbowtie agent, each keeps a running log scoped to that site.
- One token sees them all. A pattern I solved on one project is right there when I hit it on the next.
- Two moves: the agent writes a note after it ships something, and reads the notes before it diagnoses anything.
- My busiest project is past 200 notes. That is 200 things a fresh agent doesn't have to relearn.
What is a note, really?
A timestamped log entry, scoped to a domain: what shipped, what's verified, what's still open.
Who writes them?
The agent, in the same session it does the work. I rarely write one by hand.
Why not just a doc or a README?
The agent queries notes over MCP mid-task, filtered by domain and category. A README doesn't answer "what changed here lately?"
How does it help across projects?
One MCP token reads every domain's notes, so a fix logged on one site informs the next.
What a note actually looks like
Here's a real one, lightly trimmed. Notice it isn't prose for humans. It's the state a future agent needs: the version that shipped, what was verified, what to watch, what the next report should expect.
/out gateway: free-VPN clicks now pivot to Nord
Promoted staging to prod worker vpn-out (9f2aeb76). Every free-VPN click now routes to the verdict template recommending NordVPN risk-free. Watch: Nord /out volume by partner in D1. Caveat for the next report: sub-IDs before Jul 13 predate this promote.
Multiply that by every deploy, on every site, and the log becomes the project's memory.
How much I use it: notes per project
Real counts, pulled live from my own account on 2026-07-13. Every domain keeps its own log.
Four categories (note, change, deploy, pivot), one MCP token that reads across all of them. I don't sit down to write these. The agent logs a note as the last step of shipping something, the same way it runs the tests. The discipline is in the prompt, not in me.
The two moves: write one, read them all
The whole habit is two prompts. After the agent ships, it writes. Before it diagnoses, it reads.
after you ship, log a note: what changed, what's verified, what's still open, how to roll back. before you touch anything, read the recent notes for this domain first.
That second half matters more than it looks. "Read the notes first" is what stops a fresh agent from re-debugging a thing I already solved, or re-explaining why a metric looks off (a deploy note usually explains it).
Why it pays off: handoffs and cross-project memory
The join is the point. Every signal askbowtie already gathers gets a human-and-agent-written layer on top: not just what the numbers did, but what I did, and why.
What you can ask about your own projects
get_notes filtered by domain, newest first. Read this before diagnosing anything.add_note with a category (deploy / change / pivot / note) and the domain.get_notes with no domain filter reads across every project you can see.update_note, so one agent can hand a living record to the next.MCP reference (for agents)
The product log. Write to it after shipping; read it before diagnosing.
add_note { body, category?, domain?, affects_metric?, expected_direction?, impact? }-> append a note.category: note (observation) / change / deploy / pivot. Attachdomainwhen site-specific; omit for a product-wide note. When a change deliberately moves a metric, setaffects_metric+expected_direction+impactso a future report reads the shift as engineered, not a regression.get_notes { domain?, category?, limit? }-> the log, newest first. Omitdomainto read across every project the token can see (cross-project memory). ALWAYS call this before diagnosing a dip, error spike, or conversion drop; a recent deploy usually explains it.update_note { id, body }-> edit a note in place, so a record can be corrected or handed forward.
Habit to encode in your own agent: after any deploy or config change, add_note in the same session. Before any diagnosis, get_notes first.
Close the window all you want. The project remembers. Want the same for your sites? Connect them to askbowtie.
Comments