Building a system that knows when to remember, when to act, and when to ask.
Meet Alfredo, the personal operating system I am building around my actual life, one capability at a time.
Memory
Keeps useful context over time and brings it back when a decision actually needs it.
What Alfredo can do now
- Persistent context across everything Alfredo does
- Human-readable knowledge in Obsidian, instead of disappearing into some mysterious database
- Separate memory for different parts of life, rather than one giant bucket of everything I have ever said
- Durable knowledge kept separate from temporary runtime state
- Context brought back into ongoing workflows, not just when I explicitly ask Alfredo to remember something
What I learned
Alfredo does not need to remember everything about me. That would be creepy and mostly useless. He needs to remember the right thing at the right time.
Under the hood
Food
Connects meal planning, what we probably have at home, recipes, feedback, shopping and the actual grocery cart.
What Alfredo can do now
- Meal planning that stays conversational until I actually approve it
- A deliberately fuzzy pantry: confirmed, probably there, or missing, because pretending Alfredo knows there are exactly three onions would be a lie
- One canonical version of each recipe, instead of slowly creating five almost-identical ones
- Feedback kept separate from the recipe, with adjustments queued for next time
- A shopping list that stays a shopping list and is never silently “improved”
- Product matching and real cart preparation in Continente Online
- Alfredo can prepare the cart. Checkout remains a human privilege
What I learned
The useful part is not planning the meal. It is remembering what happened when we actually cooked it.
Under the hood
People
Keeps the people I care about, important dates, gift ideas and unfinished things connected early enough to actually do something about them.
What Alfredo can do now
- One canonical record per person, even when I refer to them in different ways
- Important dates treated as preparation windows, not notifications that arrive when it is already too late
- Gift ideas and gift history, so I do not have the same brilliant idea twice
- Activities and open loops attached to the person they actually belong to
- Silence when there is genuinely nothing useful left to do
What I learned
Remembering someone’s birthday on the day is technically correct and practically useless. Alfredo treats it more like a tiny project with a deadline.
Under the hood
Communication
Lets Alfredo schedule and send approved messages without ever becoming confused about who is speaking or what I actually authorised.
What Alfredo can do now
- Only people who are explicitly authorised and have opted in
- Recipient, message, time and channel shown in full before I approve anything
- A scheduled outbox that stays editable and cancellable until it runs
- Alfredo always speaks as Alfredo and never pretends to be me
- If delivery is uncertain, he does not enthusiastically send the same message again
What I learned
Once Alfredo can speak to other people, consent and authorship stop being implementation details.
Under the hood
Time
Connects calendars, tasks, reminders and context without pretending they are all the same thing.
What Alfredo can do now
- Google Calendar as the source of truth for commitments, across five accounts, because I decided that one calendar was not enough in my life
- Todoist as the source of truth for concrete next actions and to create clear tasks
- Reminders left in the system that can actually remind me, rather than hiding them inside a calendar
- Obsidian holding the context and open loops that are neither an event nor a task
- Every write tied to the exact account I approved and shown in full before it happens
- Different write policies per account, from read-only to reinforced confirmation
What I learned
“Add this to my calendar” sounds simple until Alfredo has five calendars to choose from. A write is only safe when he knows exactly which one I meant.
Under the hood
Reliability
Makes Alfredo recoverable and observable now that rebuilding him from scratch would genuinely ruin my weekend.
What Alfredo can do now
- Encrypted scheduled backups with retention and verification
- Only the state I cannot reconstruct is backed up; logs and generated files can be rebuilt
- Restores tested somewhere isolated before I trust any of them
- Restore validation covering required files and database integrity
- A private health view for capacity, sync, scheduler state and backup freshness
- Sanitised status only, because a health dashboard probably should not leak secrets while explaining that everything is healthy
What I learned
Backups became important around the same time Alfredo became useful enough that losing him would be genuinely annoying.
Under the hood
Notices commitments, missing replies and useful context without creating yet another inbox for me to manage.
Built and connected across multiple accounts. Now comes the less glamorous part: seeing whether Alfredo is useful on real email without becoming another source of notifications.
What Alfredo can do now
- One logical inbox across several authorised accounts, without copying mail out of Gmail
- Read-only by design, with sending blocked at the gateway rather than politely discouraged in a prompt
- A separate triage dedicated sub-agent that reads a bounded window and cannot act, schedule or speak
- Explicit relevance rules that are dated and reversible instead of becoming permanent truths because of one email
- Reply text can be proposed, but drafting and sending are separate approved actions
Still figuring out
How do I teach Alfredo the difference between “this actually matters” and “someone used the word urgent”?
Hard boundary
Never speaks as me, and never sends, archives or labels anything.
Under the hood
Capture
One place to throw things before I have to know what they are, why they matter or where they belong.
Teaching Alfredo next
- One capture path for links, notes, voice and whatever else I throw at Alfredo
- An ideas area for things that are still half-baked, in readable Markdown
- Anything unclear stays in the inbox instead of being confidently filed in the wrong place
- Promotion into people, projects or a wishlist once the intent becomes real, linked rather than duplicated
Still figuring out
Where should something live when even I have no idea yet what I am going to do with it?
Hard boundary
Writing something down is not the same as committing to doing it. Alfredo will need to learn the difference too.
Research
Finds the small number of things genuinely worth my attention without creating another endless thing to keep up with.
Teaching Alfredo next
- A read-only radar pointed at my actual interests, projects and open loops
- A deliberately short handful of things rather than an infinite stream
- Its own bounded specialist preparing a brief instead of taking action
- Nothing subscribed to, shared or acted on without approval
Still figuring out
I do not need Alfredo to build me another feed. The internet has already produced enough of those. I want a few things that make me think “good catch”, not 47 things I now feel guilty for not reading.
What I’m building toward
The goal is not to keep adding integrations. I want Alfredo to gradually understand enough context that it can remove some of the mental overhead from everyday life. More importantly, it needs judgement: knowing when to act, when to ask first, and when the correct action is to do absolutely nothing.