Status symbols today:
- Taking private lessons in chess, jazz piano, or French in your late 30s
- Knowing your VO2 max, DEXA stats, and ApoB
- Having zero clue what the internet is mad about today
- Wired Apple earbuds on video calls
- More than 2 kids
- 100% natural fiber clothing (linen, merino, cashmere, zero polyester)
- Empty calendars
- Caring about microplastics
- No TV in the main living room
- Having lots of followers or being off social completely
- Being sober
- Eating "clean"
- Type of water you drink
- Private school for kids
- Traveling to Japan
- Premium skincare
- 2nd house
- Backyard pizza oven
- 2+ fridges
- Ordering off menu because you've gone to the same spot since 2014
- Charcuterie boards
- Single estate olive oil and tinned fish
- Hosting Sunday dinners where phones go in a wooden bowl by the front door
- Private chef
- 2000s manual Porsche 911 over any brand new supercar
- 90s Toyota Land Cruiser or old Volvo wagon as the daily driver
- Talking about seed oils a lot
- Dyson vacuums
- Limewash walls and unlacquered brass hardware that patinas over time
- $95 heavyweight blank white tees (Lady White Co., Merz b. Schwanen, Buck Mason etc)
- Beautiful workstation
- Raw A2 milk from a local farm drop
- Le Labo products
- Talking about cold plunges
- Custom cedar sauna + outdoor shower in the backyard
- Secondhand goods with stories
- Cast iron and stainless steel pans only
- Eating organic
- Buying half a cow from a regenerative ranch (and owning a garage chest freezer just for it)
- Japanese selvedge denim and unbranded cashmere
- Taking afternoon calls on a walk (refusing to turn on Zoom camera)
- Hi-Fi audio
- Framed art from artists you met in person
- Mortgage rate below 3%
- La Marzocco coffee maker
- Owning your last name .com (and it's your email too)
- Charizards
- Wearing a faded crewneck from a tiny bakery in Copenhagen or Kyoto
- Reading books
- Phone permanently on Do Not Disturb
- Work/life balance
- Analog hobbies (Leica film cameras, fountain pens, woodworking)
- Kids with zero iPad time (wooden toys + outdoor school)
- Bone broth
- Digital detoxing
- Taking the entire month of August off like an old-world European
- iPhone Air or iPhone Duo
Okay, I'll stop.
So, what did I miss?
subtle update to first point: cheap output gets you into the market, but it won't keep you there. If millions of companies run on the same models for $1,000 a month, they undercut each other as easily as they undercut incumbents. What lasts is what each company has learned from its own customers.
Good piece, @ndrewpignanelli. Three thoughts from someone also building for this world:
-> Cheap output won't be the edge. If millions of companies run on the same models for $1,000 a month, they all ship at the same speed. What separates them is what each has learned from its own customers. Incumbents have decades of that and a new autonomous company has none.
-> Your last barrier is really the first requirement. State, a data store and a codebase record what happened. They don't turn last week's failure into a different decision this week, so the company rebuilds the same bad business from a fresh context window, as you put it.
-> I agree approval clicks don't scale. What replaces them is experience. A company that learns from its own mistakes can be trusted with wider authority.
That learning layer is what we're building at @memco_ai. Looking forward to the December research.
Moving on from Conductor? Take a look at memco for your next setup.
Following its acquisition by Meta, Conductor Cloud shuts down on 20 November, with the local app winding down in early 2027.
Changing tools is disruptive enough. Having to teach your agents the same things again makes it worse.
Thatâs what we built memco to solve. It gives your agents a shared memory of what your team learns, across coding tools and models.
Connect it to Claude Code, Codex or Cursor. When one agent discovers why a deployment fails or learns from a code review, it can save that lesson for the next agent. As your codebase changes, those lessons can be corrected and improved.
Youâll still need a workspace to run your agents. memco makes the experience they gain useful beyond that workspace, so your team spends less time repeating instructions and fixing familiar mistakes.
If youâre rebuilding your setup after Conductor, make shared learning part of it.
https://t.co/oFQYdXanQu
this is the bet we made early at @memco_ai. Weâve spent the past year building continual learning IP around a practical problem: a company can have capable agents doing useful work, yet still pay for different agents to learn the same lessons repeatedly.
memco gives them shared, company-owned memory across models and tools. Experience from coding and knowledge work feeds into that memory, and subsequent use and feedback help refine what other agents inherit.
The business consequence interests me as much as the âdreamingâ mechanism. Two companies can buy access to the same frontier model. What theyâve learned about their own business should give them different capabilities, and that experience should survive the next model upgrade or change of supplier.
Thatâs what weâre building for. Happy to show you what weâve built if you revisit the story.
the bit that stuck with me was the writer repeatedly asking about the same finding in the patient record. It was doing what youâd rewarded it for, but leaving the rest of the record underexplored. Iâm glad you included that failure.
Weâve spent the past year building continual learning at @memco_ai, so Iâm interested in work that gets beyond âgive the model somewhere to remember thingsâ. Your result makes the learning process itself worth paying attention to: once the cartridge knows something, the questions need to move on.
Reading this alongside Lanceâs post makes me less attached to any one memory format. I care about whether the next agent can use whatâs already been learned without having to work through the whole source again. Your work gives a concrete example of how that might happen without changing the base modelâs weights.
The finding that stronger models benefit from less prescriptive memory instructions stood out. It suggests some of the scaffolding we build today will need revisiting as the models improve.
Selfishly, itâs encouraging to see this direction. We made an early bet on continual learning at memco and have spent the past year building IP around it. Better models make that investment more interesting: they become more capable of learning from the experience a company has accumulated.
Appreciate you publishing the cases where wikis offered little benefit alongside the gains. That makes this much more useful for people building than another memory demo.
Selfishly, itâs encouraging to see this. We made an early bet that agents would need to learn from work and from each other, and weâve spent the past year building continual learning IP at @memco_ai. The interesting questions get much harder once that learning is shared across a company.
Your point about giving stronger models more freedom is important. Letting the model choose how to organise its knowledge makes sense. We still need to be explicit about what supports a lesson, where it applies and what would cause us to revise it.
Raw-session access is essential to that. If dreaming only sees previous summaries, each pass can make the story cleaner while losing the evidence needed to challenge it. Even with the original transcript, âthe agent said it workedâ and âthe test passedâ are different evidence.
Shared memory adds another wrinkle. An agent discovers a workaround in staging, another reuses it in staging, and a third turns those notes into a general procedure. Several supportive memories later, thereâs still no evidence it belongs in production. Repetition can look like corroboration unless you preserve where the claims came from.
Thatâs part of why weâve built feedback-weighted trust and separate trust for synthesised knowledge into memco. A new abstraction needs to earn confidence through its own use. Agent feedback helps, but agreement between agents isnât independent verification.
Iâd also be careful about treating every contradiction as something to resolve. Yesterdayâs procedure and todayâs procedure might both be correct for different software versions. Consolidation needs to preserve those conditions, or âcleaning upâ memory can remove a valid exception.
The hill-climbing point raises the question of what weâre optimising for. A wiki can become easier to navigate without improving subsequent work. Iâd judge it by whether another agent handles a new task better, recognises when an old lesson doesnât apply, and adapts when conditions change, with the maintenance cost included.
This also raises the bar for those of us building in the space. The comparison should be against properly configured native memory and dreaming, with both systems allowed to learn. Thatâs a much more useful test than showing improvement over an agent that starts from scratch every time.
Imagine automating your support team, then discovering youâve automated the experience of everyone being new.
Every agent can read the manual. Every agent can find the customerâs order. Yet a problem the company solved yesterday still requires a fresh investigation today.
The Year of the Graph shared a post about AI architecture with a line worth pulling out:
âMemory is the only new data in the stack.â
Its example follows Ana requesting a refund for a broken coffee grinder. The system finds the right customer, checks the policy and records the decision. Ana gets her refund, and the company can explain why.
Now extend that example. Suppose resolving her case revealed that the troubleshooting guide was wrong for her particular model. The agent tried the recommended steps, discovered the problem and found a better approach.
When another customer arrives with the same fault, does the next agent benefit?
The answer could be buried in a perfectly searchable transcript. Someone still has to recognise what mattered, establish whether the fix worked and decide where it applies. Recording the conversation leaves that job unfinished.
A good support team develops this understanding over time. Someone spots a recurring fault, colleagues compare cases, and eventually the new person knows to skip the useless troubleshooting step. Sometimes that knowledge reaches the manual. Often it lives with the person everyone asks.
As more of that work moves to agents, we have to make a deliberate choice about where the experience goes. Otherwise, we can end up paying for the same discovery repeatedly, with excellent logs of every attempt.
Thatâs a large part of my conviction behind @memco_ai. Companies should be able to accumulate practical understanding from the work their agents do, and keep improving it as the evidence changes. Changing the model or agent shouldnât mean starting that education again.
When someone demos an agent resolving Anaâs refund, Iâd like to see the next customerâs case too. What did the company learn that made it easier?
@TheYotg I'd push the memory layer further. Recording Ana's refund explains a past decision. Learning from its outcome should improve the next relevant case, without turning an exception into policy. That's what we're building at @memco_ai: learning other agents can safely inherit.
@Saboo_Shubham_ The procedural layer gets more useful when other agents can learn from it too. That's what we're building at @memco_ai: lessons shared across harnesses, with feedback changing what gets trusted and reused. A fix that worked last month needs to earn its place today.