The Avatar That Remembers the City
Mebots, self-modeling, and real-world memory as game state
A Mebot can be treated as a self-modeling avatar: a figure inside the city through which the player’s real-world behavior becomes game-state and memory.
That sounds like game design, but it points at a deeper cognitive pattern. A mind does not only perceive the world. It also builds a model of itself inside the world: a body, a position, a set of capabilities, a history, a felt sense of what it can do next. The self is not the whole system. It is the handle the system uses to act.
A Mebot can play that role for a city game.
The player walks, visits, patrols, scouts, claims portals, returns to favorite routes, avoids certain zones, repairs damaged habitats, and learns which places matter. The Mebot is the compact visible agent that absorbs those actions. It is not just a character with stats. It is the game’s model of the player’s situated agency in the city.
The map remembers through the avatar.
The avatar is a handle for action
An avatar usually means representation. A little figure stands in for the player. It has cosmetics, levels, maybe inventory. That is the thin version.
The stronger version is functional. The avatar is how the system turns a messy human life into an action-ready model. It does not need to know everything about the player. It needs to know enough to answer practical questions:
- what territory does this player protect?
- what kinds of places do they repeatedly choose?
- which routes are familiar?
- which enemy habitats have they sensed?
- what kinds of rewards actually move them?
- what does their city look like from their pattern of action?
This is self-modeling in a modest, designable sense. The Mebot is not conscious. It is not the player. It is a structured proxy: a playable body that carries the player’s local history forward.
When the player moves, the avatar changes. When the avatar changes, the map can suggest new moves. The loop is what matters.
Movement becomes memory
Most location apps treat movement as a coordinate stream. That is useful for routing, but cognitively thin. Coordinates by themselves do not say what the movement meant.
A Mebot-style loop can give movement a role.
Visiting an owned habitat becomes patrol. Visiting an enemy habitat becomes scouting. Returning to the same cafe becomes coffee memory. Holding a Business Portal becomes guardianship. Walking the same path repeatedly becomes a corridor. A raid leaves a wound. A repair visit becomes regeneration.
The raw substrate is still ordinary city activity. The cognitive layer is the interpretation.
This is where the game matters. A game can turn action into state without requiring the player to fill out forms. It can make memory visible as color, scent, scars, badges, routes, and changing affordances. The player does not need a dashboard explaining their life. They need a living map that says: this place is protected, this border was scouted, this portal matters to you, this route is becoming yours.
Memory should not mean surveillance. The important object is not a private GPS trace. It is a derived signal that can decay, aggregate, and become useful without exposing the person underneath it.
The city as a model of the player
A city map usually models streets and places. A game map can model relations.
For one player, a cafe is a morning anchor. For another, it is irrelevant. A park may be a recovery node, a gym may be a stamina organ, a lunch place may be a social waypoint, a transit stop may be an artery. The objective city is the same. The lived city is different.
The Mebot becomes the interface between those two maps.
It carries the player’s revealed city: the places they return to, the distances they tolerate, the categories they care about, the habitats they protect, the risks they take, the routes they repeat. Over time, the Mebot can earn category identities: Coffee Scout, Lunch Pathfinder, Work Spot Cartographer, Patrol Keeper, Portal Guardian.
Those titles are not just rewards. They are compressed memories. They say: this player has produced enough verified behavior in this slice of the city that the system can start treating them as locally competent there.
A title is a cognitive shortcut.
Patrol and scout as perception-action loops
The recent patrol/scout mechanic is important because it turns the map from display into feedback.
A feedback loop has a simple structure: sense, act, update, act again. If a player visits an owned habitat, the city records a protective signal. The map shows a patrol field. The field fades. The fading creates pressure to return. If the player visits enemy territory, the city records a scout signal. The map shows fresh sensory memory. The memory ages. The player chooses whether to raid, avoid, or scout again.
The game state is not a decoration on top of movement. It is what movement becomes once the system understands it.
This is where a Mebot starts to feel less like an avatar and more like a small cognitive body. It has a perimeter, memories, wounds, preferences, and unfinished business. It is pushed around by gradients: fading patrols, fresh scout scent, contested portals, damaged habitats, category progress.
The player’s body moves through the real city. The Mebot’s body moves through the remembered city.
Grounding for agents
AI systems are often weakly coupled to reality. They can generate fluent maps of the world from text, but they do not automatically know what is open, what is trusted, what was actually visited, what route felt good, what place became useful, or which local preference survived contact with reality.
A Mebot loop can produce a different kind of data.
Not passive location data. Not generic reviews. Not scraped listings. Human-verified, game-contextual signals:
- this place was visited;
- this portal was claimed;
- this category attracted repeat behavior;
- this route became a habit;
- this area was scouted before action;
- this habitat was defended before decay;
- this place mattered enough to return to.
For an AI agent, those signals are valuable because they are grounded in action. They bind language and map data to human behavior in the world.
The Mebot is the compression layer. It turns many small acts into a local model an agent can use.
The avatar should not become a dossier
There is a failure mode here. A self-modeling avatar can become creepy if the system treats the player as an object to profile rather than an agent to empower.
The design line is clear. The Mebot should expose game memory, not surveillance memory.
Good memory sounds like:
- Coffee Scout memory grew.
- Patrol field refreshed.
- Work Spot confirmed.
- This route is becoming familiar.
- Scout scent fading on the east border.
Bad memory sounds like the system is watching a person too closely. Exact traces, private routines, timestamps, and inferred personal facts should not become the product surface.
The avatar should make the player feel more capable in the city, not more inspected by the software.
This is also better cognition design. A useful self-model is selective. It does not contain everything. It contains what helps the system act.
A playable self-model
The concise statement is this:
A Mebot is a playable self-model of the player’s city behavior. It is the place where real movement turns into memory, where memory turns into map state, and where map state turns into the next possible action.
That makes the avatar more than decoration. It is a handoff between substrates.
The player acts in streets, cafes, parks, gyms, offices, borders, and routes. The game receives those acts as visits, claims, patrols, scouts, scars, titles, and portal memories. The AI layer can then reason over the compressed pattern without needing to own the raw life underneath it.
The player gets a living city-body. The Mebot gets a memory. The map gets a self.
Source conversation: Future Day 2026: Joscha Bach, Anders Sandberg, Adam Ford.