What should no longer govern what happens next?

I'm Michael. I live and work in SoMa and the Tenderloin, and I build spatial and agent systems that try to keep people and places in view. Yesterday I spent the day at the Jevathon, a hackathon built around TypeSafe's JEV decision model, and I kept experimenting into this morning.

Something small came out of those runs that I didn't expect, and it lines up closely with what this gathering is asking about. I wrote it up so I'd have something concrete to hand people today, and so the conversation can keep going after the afternoon ends.

I was testing how memory changes the way agents play simple Atari-style games. Two things happened.

First, giving the fast agent more memory didn't help it. Its results stayed about where they were, and each decision took far longer because it had more to sift through before acting.

Second, in one condition I let a smaller model inherit a conclusion from an earlier round of experiments. The conditions had changed since that round. The model treated the old conclusion as a description of the present, and its performance fell apart.

On the technical side this is ordinary. Engineers call it a stale prior. What stayed with me is how familiar it felt from the human side. I've done exactly this. Most of us have, in relationships, in organizations, and in the way we walk past a block we think we already understand.

These were small, informal runs over a weekend. I'm offering them as a prompt for conversation. They don't prove anything about wisdom.

Related hexagrams:

Most current work on AI memory treats forgetting as a defect to overcome. The push is toward longer context windows, persistent memory, perfect retrieval, complete histories and knowledge graphs that only grow. The implied goal is a system that loses nothing.

Many wisdom traditions carry a warning pointed in the other direction: the residue of the past can bind us. In Indian thought, saṃskāras are the impressions left by past action that shape how we perceive and respond, often below awareness. Ṛṇānubandha names the bonds of debt formed through past interactions that keep drawing beings back into relation with each other.

I'm holding these as analogies. I don't think the old texts anticipated vector databases. What they offer is a long-tested description of what happens when the past keeps acting on the present, which is the same problem my agents ran into on a much smaller scale.

Related hexagrams: ,

Tyson Yunkaporta, an Aboriginal scholar of the Apalech Clan in far north Queensland, wrote Sand Talk as a way of thinking through global problems with Indigenous pattern thinking. He draws symbols in the sand as he yarns, and the drawings carry the argument as much as the words.

A few ideas from the book keep coming back to me in this work. As I read him, knowledge lives in relationships: between people, between people and land, and across generations. It thins out when it gets pulled out of those relationships and stored as abstraction. On that view intelligence belongs to a network of relations more than to a single head, and there are many kinds of it, including the kind a landscape holds.

He also describes a sequence for how knowledge moves through a group: respect, connect, reflect, direct. Each step depends on the ones before it, so nobody gets to direct until they've done the first three.

That reads differently after watching my agents. A swarm that shares distilled findings is a crude network of relations. The four steps look a lot like a protocol for deciding when a new finding has earned the right to steer the group. The agent that inherited yesterday's conclusion and acted on it right away had skipped reflect entirely.

He is also sharp about what happens when a system forgets that it is one part of a larger pattern and starts treating itself as the center. I take that as a warning aimed at my own field.

I'm a newcomer to this material, and it comes from living cultures with their own protocols about who carries what knowledge. I'm sharing how it has shifted my thinking. Please read the book itself.

Related hexagrams:

The working definition I've landed on: memory is the persistence of the past in a way that changes what's possible in the present.

By that definition memory reaches well past storage. A habit counts, and so do a scar, a path worn through grass, an office procedure, an ontology, a relationship and a culture. So does the salience landscape an agent learns, its sense of what's worth looking at. A city counts too. The Tenderloin carries decades of decisions about who receives care and who gets walked past, and those decisions shape what seems possible on any given block today.

When a group of agents inherits findings from agents it never met, it has something like what Alfred Korzybski called time-binding. That brings me to the question I keep returning to: when does memory become wisdom, and when does it become inertia?

From my runs, a rough ladder:

No memory

The agent rediscovers everything, every time.

Useful memory

Learning accumulates.

Memory held too firmly

Lock-in. Yesterday's answer overrides today's evidence.

Too much retrieval

Decisions get slower and heavier without getting better.

Context-sensitive memory

Possibly where wisdom starts.

Related hexagrams: ,

Korzybski is also the source of the line "the map is not the territory." I think about it every time I finish a Gaussian splat scan of a room or a block. A good scan is useful and often beautiful. It is still a representation, and the place keeps going after the capture ends. The light moves, people pass through, the corner smells different at night.

A lot of what a place knows is quiet. It shows up in the doorway people route around, in where the elders sit, in the bench that always stays empty, in the plants that come back after every cleanup. You learn it by staying and listening, usually over a long time and usually with the people who have been there longest. In that sense wisdom sits in places, and it rarely announces itself.

This matters for AI because models are representations of representations. An agent reading a dataset about the Tenderloin is holding a map of a map. The risk I worry about most is quiet substitution: the representation becomes what decision-makers consult, and the place itself stops being asked.

I've been calling my response to this memory as a boundary object. A shared memory of a place, built this way, would keep several perspectives, disagreements and open unknowns side by side, and would point people back to the place to go and listen. It should resist letting any one observer's picture quietly become the place.

Related hexagrams: , ,

Three terms I've been working with, and the relationship between them is the part I most want to talk about today.

My working guess is that wisdom lives in how these three relate. A wise person remembers without being imprisoned by memory, pays attention without being captured by whatever is loudest, and acts without pretending to a certainty they don't have.

That's a philosophical proposal. The experiments don't prove it. They do give me a concrete system where I can poke at it and see what happens.

Jung's shadow is the part of ourselves we don't acknowledge: motives we'd rather not see, such as the wish for control, recognition, status or certainty, along with projection, rationalization and displaced aggression. AI safety covers a lot of the same ground in different vocabulary, under names like power-seeking and specification gaming. Individuation raises the chance that a person can notice "part of me wants control here" before that wish gets written into an institution, a metric, a dataset or an agent.

A capable system doesn't need a shadow of its own for ours to act through it. The path looks roughly like this:

1. Unconscious motive
2. ›
3. Stated objective
4. ›
5. Proxy metric
6. ›
7. Automated optimization
8. ›
9. Institutional scale

Groups have shadows too. Organizations bury inconvenient information, cultures build out-groups, and movements persuade themselves that their own power is uniquely righteous. AI can turn those distortions into very efficient procedures, which is why collective shadow work belongs in the safety conversation.

I'd stop well short of reducing safety to spiritual maturity. We still need evals, interpretability, security, governance, adversarial testing, corrigibility and checks on concentrated power. Inner work improves the people who design and govern those mechanisms. The mechanisms still have to exist.

The thread running through all of this is partiality. A partial observer is safer when it knows it's partial. Much of individuation is learning that my model of reality is a model. Memory as a boundary object tries to do the same thing for a group, and the place work above tries to do it for a city.

Jung didn't think the shadow could be removed, and I don't think we should design as though it could. The principle I'd propose instead is to build people, institutions and AI systems that can meet what they've excluded without immediately projecting it, suppressing it or optimizing it away. Unpacked, that turns into fairly concrete requirements:

• Keep dissenting and minority findings visible in shared memory instead of averaging them out.
• Record what an objective leaves out, and who decided to leave it out.
• Send anomalies and contradictions to people instead of quietly smoothing them over.
• Let the objective itself be questioned during operation, and make that easy.

Related hexagrams: ,

This is one of the questions on today's invitation, and my suspicion is that trying to formalize wisdom itself is where things go wrong. Humility probably won't survive being turned into a number between 0 and 1.

We can formalize some of the conditions that make wiser behavior more likely. Those include keeping uncertainty visible, allowing beliefs to be revised, tracking where each memory came from (its provenance), letting things be forgotten, knowing when to stop, and noticing when the context has changed.

Each of those can be tested. The simplest test I've found asks whether a system changes its mind when reality contradicts it. Contemplative practice and engineering both have ways of checking that, which makes it a good place for the groups in this room to meet.

For a second hackathon this week, I'm setting up a small swarm. Somewhere between 8 and 16 worker agents each try strategies in parallel game environments running on Tenki. They share only distilled findings through a common memory layer, Mitosis Cortex, and never full transcripts. JEV looks at the state of the whole group and decides who acts next, what deserves another try, what gets dropped, and when there's enough evidence to stop.

Halfway through, I change the rules so that a strategy that had been winning stops working. This is a regime change. Then I compare two groups. In the first, memory accumulates and old findings stay just as authoritative as new ones. In the second, memory is revisable: each finding carries where and when it came from, and newer evidence can weaken or replace it.

I'll track games until discovery, redundant experiments, the decisions JEV makes, compute cost, time to converge, and above all how long each group takes to adapt after the change. That adaptation lag is the number I care about most. I'll post results back here.

Related hexagrams: ,

The invitation also asks what practices help the people building this technology stay whole. My answer is modest. I think of it as service as ritual: small, consistent acts of care, done together, for ourselves, for each other, and for neighbors who are struggling.

For me that currently means the Tenderloin street cleanups run by Refuse Refuse SF, where volunteers are thanked in SFLuv, a local currency spent at neighborhood merchants. It gets my attention off a screen and keeps the people I'm building for in front of me. It is also a way of listening to a place. Showing up again and again is how a group wears in a path it actually wants to keep.

I'd love for Apollo to adopt one service practice like this as part of who we are. If that interests you, find me today.

Related hexagrams:

If intelligence asks what it knows, and memory asks what should persist, maybe wisdom starts by asking what from the past should no longer govern what happens next.

• In your tradition, what practices help someone let go of a conclusion that used to be true?
• What does a place know that no scan or dataset of it can hold, and how do you listen for it?
• How does a community decide which of its memories stay authoritative, and who gets to make that call?
• What would it look like to do shadow work on an objective function before it ships?
• Where have you seen more information make a person, or a system, less wise?