One conversation I wanted to write down from SF Tech Week started with coconut.
What’s your earliest memory of coconut?
Heritage Table at SF Tech Week
Scarlett and I were in San Francisco during SF Tech Week, October 5–11, 2026. Presented by a16z, it brings together independently hosted events across the city, including founder gatherings, investor meetups, product discussions, and community dinners.
We were there to introduce Okou, meet people working in different fields, and hear how they actually use AI: what work they hand over, how they teach an agent, and what makes them trust it enough to keep using it.
One of the gatherings we supported was Heritage Table’s “The Memory Table: What Should AI Remember?” On October 9, we joined hosts Evangeline Medina and Pubali Sen at artist Peggy Gyulai’s loft for a shared dinner and conversation about food, memory, and technology.
Heritage Table helps families preserve recipes alongside the voices, stories, and people behind them. The Memory Table brings that idea into a shared meal. Its SF Tech Week edition gathered founders, investors, AI builders, and thinkers around one long table.
That’s where the coconut question came in.
Coconut was the thread running through the meal—an ingredient that carries different memories across families and cultures. Guests shared their earliest coconut memories: a dish, a person, a place. Those stories opened up a conversation about what it means to preserve something, what context needs to stay attached to it, and who gets to decide what is remembered.
These questions felt relevant to what we’re building at Okou. When someone teaches an agent how they work, they are giving it context, preferences, and judgment. They also need to understand what it will retain and who else can use that information.
The dinner made that question tangible. Conversations were not recorded by default. Guests could choose what to contribute to a shared Living Cookbook, whether a recipe, a memory, a photo, or a few words about the evening. Preserving a memory was a deliberate choice.
The Memory Table gathering, October 9. Photo shared by Evangeline Medina.
These are six reflections from the conversations, connected with what we have been discussing internally at Okou.
1. People expect to teach an agent through working together.
The expectation is straightforward: we do something together, I give you feedback, and next time you understand how I want it done.
I’ve seen this in our recent discussions, too. Even people paying for several AI products struggle to use “skills.” They describe the experience through teaching and memory. They want to know whether the agent remembers the work they just completed together.
For Okou, we need to make that progression understandable: clarify the task, produce something, review it, and explain what will carry forward. People should see that their corrections made a difference.
2. An agent’s role needs concrete responsibilities.
A title such as “marketing agent” creates expectations without explaining what the agent owns. What does it deliver? What counts as good work? Which decisions can it make?
A useful role starts with responsibilities: research prospects, prepare a brief, draft a follow-up, and identify where review is needed. The human’s role also becomes clearer: set direction, explain standards, and decide how much authority to delegate.
This is an onboarding problem as much as a capability problem.
3. “The agent is a snapshot; the human is an amplifier.”
An agent carries a version of the context, standards, and working methods we have given it. That snapshot can change as it receives new information and feedback.
People expand what goes into it. They bring experience from elsewhere, question an assumption, build a relationship, or recognise that the business needs a different direction.
The opportunity is to help that judgment travel further. Something one person figures out can become a method the team reuses—and revises as it learns.
4. People already talk about “hiring” and “firing” agents.
They give an agent a job, invest time in explaining it, judge its performance, and replace it when it stops being useful.
That makes recovery important. When something goes wrong, can the user tell whether the problem was an unclear task, missing context, an outdated instruction, or insufficient capability?
If starting over is the only practical option, the time spent teaching the agent feels lost. We should help people correct the working relationship and understand what improved.
5. Memory creates questions about ownership and permission.
The dinner’s invitation to choose what enters the Living Cookbook connects directly to this: remembering something and making it available to others are separate decisions.
At Okou, we need to distinguish personal memory, team instructions and workflows, and connected context. Calling all of this “team memory” would misrepresent what we currently offer.
A personal correction should not silently become a rule for everyone. Sharing a workflow should not imply sharing its creator’s private information or access.
People need to understand what was retained, who can read or change it, and what happens when access to the original source changes. Those boundaries affect whether someone feels comfortable teaching an agent in the first place.
6. Founders who initially wanted to avoid hiring are starting to hire.
Another observation was that some founders who began by trying to do everything themselves with agents are now building teams.
My interpretation is that agents help founders carry more work and postpone some hiring decisions. As the company develops, they still need people who can take ownership, bring expertise, build relationships, and expand its thinking.
This connects back to the snapshot and amplifier idea. Okou needs to support both stages: helping a founder get started, then helping a growing team reuse effective ways of working while keeping personal context and permissions clear.
What I brought back
I came back with concrete things to work on: make teaching an agent easier to understand, make its responsibilities clearer, and make sharing useful work possible without blurring personal boundaries. Those are practical parts of making AI work multiplayer.
The coconut question stayed with me, too. A memory comes with a person, a story, and a choice about sharing it. The dinner gave us a way to talk about all three.
Thank you to Evangeline, Pubali, Peggy, and everyone around the table for making room for this conversation.
More about The Memory Table and SF Tech Week.
