Writing Decisions for a Future AI Assistant
I treat my Obsidian vault as long-term memory for my AI tools. When I write decisions down, an assistant can reconstruct what I did weeks later.
The problem: my tools forget faster than I do
Most AI tools are smart for a short stretch of time. After that, the context window closes, the chat scrolls away, and anything you did last month might as well have happened in another universe.
But the work you ask them to help with is rarely that short-lived. You come back to a project weeks later and need to remember:
- Why you picked one approach over another
- What you already tried and ruled out
- Where the half-finished ideas live
AI is very good at searching and summarising text. It is very bad at guessing what you were thinking if you never wrote it down. So I treat my notes vault as the memory my tools are allowed to use.
My vault is the memory, the AI is just the reader
I keep project notes in Obsidian. The important part is not the app. The important part is that I have a place where:
- Text is plain and portable
- Files are mine, local first
- I can decide which folders an assistant is allowed to read
When an AI assistant gets read access to that vault, something simple happens: it can look up how something was set up weeks ago, as long as I wrote it down at the time.
That changes what I bother to record. I am no longer writing only for “future me” with a foggy memory. I am writing for a future assistant that can:
- Search across projects instantly
- Summarise old decisions into a short brief
- Spot related work I forgot existed
So the question becomes: what exactly should be in those notes so an assistant can do something useful with them?
What a good log line looks like
I treat each project note or daily note like a tiny changelog. The basic structure of a good log line is:
- Done: what actually changed
- Decision: the choice I made, and why
- Open: what is still undecided or blocked
Written out, that might look like this:
2026-11-12
Done: Implemented basic user signup with email + magic link.
Decision: Skipped social login for now to reduce scope.
Reason: Only a small group of invited testers.
Open: Revisit social login if we open beta publicly.
This is short, but it gives an assistant several hooks:
I wrote about this earlier in Two Macs, One SSD, and the Sync I Stopped Trusting.
- “Implemented basic user signup” tells it what exists.
- “Skipped social login” tells it what does not exist and that this was deliberate.
- The reason gives context, so it can tell whether that decision still makes sense.
- “Revisit” signals future work.
Compare that to a vague line like “Worked on auth”. A human might remember what that meant for a few days. An assistant cannot reconstruct anything useful from it later.
Decisions matter more than steps
You can log every step you took, but most of those steps are noise. When I write for an assistant, I focus on decisions instead of blow-by-blow activity.
Steps sound like this:
- “Tried library A”
- “Read some blog posts about auth”
- “Refactored the signup form”
Decisions sound like this:
- “Chose library A over B because it supports magic links out of the box.”
- “Dropped phone-based 2FA for now; too much friction for test users.”
- “Moved validation to the server to avoid double-maintaining rules.”
When you come back later, or when an assistant reads your vault, decisions are more useful:
- They explain why the current state looks the way it does.
- They show what you already considered and rejected.
- They give criteria you can reuse for similar choices.
An assistant can then answer questions like:
- “Summarise all decisions about authentication for this app so far.”
- “What trade-offs did I make when I picked the current stack?”
- “Which open questions did I leave around billing?”
If your notes only contain steps, the assistant can tell you what you did, but not what you were thinking. That is usually the part you actually need.
What I let an assistant read (and what I keep off limits)
Letting an AI read your notes vault should not mean giving it your entire life. I treat my vault as several zones with different levels of access.
The “green zone” is material I am comfortable letting tools read and index:
- Project logs for client work, as long as they do not contain sensitive client data
- Technical notes, code snippets, configuration decisions
- Public writing drafts and outlines
The “red zone” is off limits:
More on this in my article Why I finally deleted NotePlan.
- Anything with private data about other people
- Raw journal entries, health notes, family notes
- Client secrets, API keys, internal documents that are not mine to share
You can enforce this with folder-level rules: point your assistant only at specific directories, or maintain a separate “AI-readable” vault that you periodically copy project notes into.
The filter does not need to be perfect, but it should be intentional. If I would not paste a note into a chat window manually, I do not want an assistant to scrape it automatically either.
What I still do not write down
This all sounds neat and disciplined on paper. In practice, there are still plenty of things I do not log well:
- Quick experiments that I abandon within an hour
- Micro-decisions that feel “too small” in the moment
- Context-switching chaos on busy days
Sometimes I only write a decision down after it hurts. I waste time re-learning how a system works, then finally add a note so I do not repeat it a third time.
I am also not turning my life into a quantified database. Not every thought or preference needs to be in the vault. The line I use is simple: if I expect to touch this project again and the decision was non-trivial, it probably deserves a log line.
That is still far from perfect coverage, but it is enough that an assistant can reconstruct most of my active work when I ask for help.
How this changes the way I use AI tools
Once you have a habit of logging decisions, the interaction with AI tools becomes more interesting than “chat about whatever is on screen”.
Instead of dumping a wall of context into a prompt every time, you can say:
Read also From hobby banknote app to an enterprise AI pitch.
- “Read the project log for <project> and summarise all open items.”
- “Given the decisions in this note, propose three options for the next step.”
- “Check whether this new idea conflicts with any past decisions in the vault.”
The assistant is not guessing anymore. It is reading your own history and reflecting it back at you.
It also nudges you to write in a way that is legible to both humans and machines: short, explicit, with reasons attached. Even if all the AI disappeared tomorrow, that style of logging would still be useful for you and anyone you collaborate with.
Writing for the reader that never gets tired
If you are already using AI tools for work, it helps to treat your notes as the long-term memory they cannot keep for themselves.
Write down what changed, what you decided, and what remains open. Give your future assistant something to read. It will happily re-read the same boring project log as often as needed so you do not have to.