Affect is what an interaction feels like. Valence is what a resident has learned about a target across interactions. Temperance is an attempt to build the second one, and to make every step of it inspectable.
A conversational AI will tell you it likes you. It will hold a tone, remember what was said ten minutes ago, answer warmth with warmth and irritation with care. It's convincing, and most of it is real as far as it goes.
What isn't there is underneath. The system doesn't accumulate a durable record of how particular people, places, ideas and experiences have affected it over time. Ask it about you and it reaches for what's in the window. Not for anything it has learned.
Temperance is my attempt to build that missing layer and then find out what happens when it exists.
It's an experimental system for studying whether an AI resident can develop persistent, evidence-linked associations toward specific targets — and whether those associations later change how it appraises, remembers and behaves.
It is not a sentiment model. It is not a mood simulator. And it is not an argument that a language model feels anything in the human sense. It's an experiment in continuity.
We already had an affect layer, and it wasn't enough. That layer performs a deterministic appraisal of an interaction. It reads the text being spoken, the framing, recent memory, stated preferences and boundaries, and produces values across a fixed set of affective channels. Those values are the system's appraisal of the interaction. They are not measurements of physiology, consciousness or subjective feeling, and I'd rather say that plainly than let the vocabulary do the work for me.
What that layer answers is: what is the affective character of this interaction right now?
Temperance adds a second question: what has this resident learned, over time, about this particular person, object, place or idea?
That second question is what we mean by valence.
Valence usually gets simplified into pleasant versus unpleasant. Temperance keeps that idea but pins it to specific targets — a person, a place, a recurring activity, an object, a phrase, a shared ritual. A resident can accumulate positive associations toward one target and negative associations toward another, at the same time, without either one cancelling the other. That independence is deliberate, and it's one of the few design decisions here I'd defend as a flat rule: the two channels are stored separately and never netted. A scalar would have been simpler. It would also have erased ambivalence, which is a result, not noise.
The part that matters is that those associations are learned from recorded interactions, not assigned in a character prompt. If interactions with a particular person repeatedly land well, evidence accumulates in one direction. If interactions around another target keep producing conflict, apprehension or aversion, it moves the other way. Not a global "good mood," but this specific thing has tended to produce this kind of appraisal.
I didn't want a system that declares "I like Clark: +0.8" and moves on. An association has to be tied to evidence: the observations that contributed to it, the target those observations were attributed to, the revisions it has been through, and its history over time. That means we can ask not only what the current association is, but why. If an AI is going to develop durable preferences and aversions, I want them inspectable rather than mysterious hidden state.
Affect and valence are different, and it's easy to blur them. Affect describes the appraisal of an event. A valence association describes what has been learned about a target across events. A playful exchange with someone might produce strong amusement and curiosity — and that one exchange is one piece of evidence toward the association with that person. One good afternoon shouldn't establish a permanent relationship. Repeated evidence can.
Two timescales, then. What is happening to the resident's affective state now, and what has the resident learned about this target over time. Keeping those separate is one of the more important pieces of the architecture.
One detail worth being exact about: the affect channels aren't signed. They're nonnegative magnitudes. Fear isn't "negative fear." Curiosity isn't "positive curiosity." They're activation levels, and a state might read conceptually as curiosity 0.7, amusement 0.5, apprehension 0.2, irritation 0.0. Valence is a separate reading of whether the relationship between the resident and a target has become broadly attractive, aversive or ambivalent.
That distinction isn't pedantry. It's required arithmetic the moment you try to measure the shape of an affective state, and that's where entropy comes in.
Entropy is defined over a probability distribution, so every term in it has to be a nonnegative magnitude:
pi ≥ 0, Σi pi = 1
If a resident's state reads curiosity 0.40, amusement 0.30, attraction 0.20, apprehension 0.10 — an illustration on a wider vocabulary than Temperance currently uses — those are all intensities. Fear at 0.4 is +0.4, not −0.4. The word "negative" in "negative emotion" isn't an arithmetic sign — it's a judgment about whether you want the thing. The channel carries activation, not that judgment, and it shouldn't carry it. If the values were signed — joy at +0.7, fear at −0.5 — you could not feed them into Shannon entropy at all. You'd have to take magnitudes first, which means keeping the magnitudes honest in the first place is the simpler design.
Given the magnitudes a = (a1 … an), normalize them:
pi = ai / Σj aj
and compute
H = Σi pi · log(1 / pi)
which is the same quantity as the more familiar −Σ pᵢ log pᵢ, because log(1/pᵢ) is nonnegative for any 0 < pᵢ ≤ 1. I write it in the first form for a reason: nothing in the formula then looks like a minus sign that a reader could mistake for a negative emotion.
Worked, on a four-channel illustration: a = (4, 3, 2, 1) → p = (.4, .3, .2, .1) → H ≈ 1.85 bits. The distribution is uneven, so the entropy sits below the four-channel maximum — but only slightly, at about 92% of the ceiling. A state of (1, 1, 1, 1) is 2.0 bits: the same total activation, spread evenly, fully diffuse. More generally, only the uniform distribution reaches the maximum — that vector is one way to be uniform, not the only one.
One caution, because it bit me: raw bits are not comparable across a different number of channels. For n channels the maximum is log₂ n. Four channels can reach 2 bits, and (4, 3, 2, 1) sits at 92% of that ceiling. Give that same state a twelve-channel vocabulary — the one used by Temperance's predecessor line, with the other eight channels at zero — and the entropy is unchanged, because the zeros contribute nothing, but the ceiling becomes log₂ 12 ≈ 3.58 bits, and the state is now at about 52% of what that vocabulary can express. So I keep two numbers: the raw bits when comparing one vector against itself over time, and the fraction of the ceiling when comparing states across vocabularies. Reading a raw bit count as "emotional complexity" without knowing the ceiling is how a low-dimensional state gets flattered.
And I should be candid about the arity, because it bears on how much of this the system currently does. Temperance's own appraisal emits two signals, not four and not twelve: a positive activation and a negative activation, independently. Entropy over a two-channel vector is a poor measurement — two values normalize to a distribution with very little room to vary, and the ceiling is one bit. So the four-channel example above is an illustration of the measure, not a description of what Temperance currently produces. Entropy over an affect vector is a proposed measurement: it becomes applicable if the appraisal layer widens to something like the predecessor line's channel vocabulary, and it would then be a measurement of that vocabulary's state, not of Temperance's. Keeping the arithmetic honest about which layer it applies to is the whole point of writing it down.
This is not "emotional chaos." It's a measure of distribution — diversity of activation, not pleasantness and not intensity. And it gives us a set of genuinely different axes: intensity, valence, affective diversity, temporal stability. Those are not one property wearing four hats.
The learning half is also just arithmetic, and it's the arithmetic I'd point at if you want to see the difference between a mood and a history.
Each channel of each association carries two numbers. One is acquired strength, `g`. The other is suppression, `x`. What gets expressed is the product of the two:
expressed = g · (1 − x)
Reinforcement moves them like this:
g′ = g + α · d · (1 − g)
x′ = x · (1 − ρ · d)
with α and ρ at 0.20 in the baseline, `d` the dose admitted for this observation. That first equation is the whole argument. For full doses from independent evidence, acquired strength after n reinforcements is
gn = 1 − (1 − α)n
which runs 0.20, 0.36, 0.49, 0.59, 0.67, 0.74.
The first full qualifying dose moves it a fifth of the way. The sixth independent qualifying dose moves it less than a tenth of what remains — and those doses have to arrive as separate evidence families, which the constraint below makes deliberately expensive. It approaches the ceiling and never arrives, and there is no value of n at which one enjoyable afternoon becomes a permanent relationship. The shape of that curve is doing the moral work that a paragraph about "gradual learning" would only assert.
Suppression runs the other way, and it's where the model gets to say something I care about:
x′ = x + β · d · e · (1 − x)
β = 0.20, `d` the dose, `e` the expectation that the outcome would occur. An omission raises suppression without erasing acquisition. Expressed strength falls; what was learned stays where it got to. Which means an association can weaken without being forgotten — that was one of the questions I most wanted an answer to.
But silence is not evidence. An omission only counts if we scheduled the trial beforehand and later established, from a captured event, that the expected thing did not occur. You cannot infer an omission from a quiet room. And elapsed time does not do this at all. We deliberately did not implement clock decay as extinction, because a system that weakens an association merely because a calendar moved is fabricating evidence. Extinction has to be earned by something that didn't happen.
There's one more constraint, and it's the one that stops a good week from hardening into a bond. Evidence is counted by observation family, and each family can contribute at most a total dose of 1.0. Repetition inside the same family does not stack. Consolidation into the slow disposition requires at least three fresh families and at least two distinct episodes before it applies at all — and when it does, it moves slowly:
v′ = v + λ · (E − v)
λ = 0.05, with E the mean expressed strength of that batch of families. Five percent per eligible batch. Reach two-thirds of the way to a stable value and you've spent about twenty consistent batches. So the fast channel answers how does this look right now, and the slow one is what the resident would eventually be entitled to call a disposition — and it is deliberately hard to hurry.
The arithmetic is decimal, twelve places, half-even rounding, so replaying the same evidence reproduces the same number exactly rather than approximately. That matters more than it sounds: an association you can't recompute is an association you can't audit.
I should be precise about the status of all of this. These constants are the baseline — an engineering comparator, not a claim about biology. The arithmetic is tested; the behavioral half is not yet demonstrated. The machinery for learned associations exists and runs, and the honest description of its live history is unexercised. Built is not the same as shown, and the difference is the entire discipline of this project. The claim-by-claim status of this work is kept in the claims ledger, where the central thesis still sits as an open question.
The discovery that surprised me was about memory. A resident can't develop meaningful long-term associations without meaningful long-term memory. That sounds obvious, and it is. But during development we found the system preserving raw conversation perfectly while ordinary interactions were not becoming durable episodic memories. The evidence existed. The experience did not. A transcript sitting in a database is not the same thing as a memory available to the resident, and the gap between those two is easy to live with for months without noticing.
We traced the missing link and added an automatic episode-write path. After a successful interaction, Temperance deterministically creates a recorded experience from the user event, any causally linked tool events, and the resident's response. No model is asked to summarize it. No model is asked to decide whether the event was important enough to remember. The interaction itself becomes part of the resident's durable experience.
That is the difference between experience → durable episode → selective recall → later behavior and experience → transcript → eventually falls out of context. I think the second one is what most systems actually do. Closing that gap turned out to be foundational rather than incidental — and note that it's the same design instinct as the arithmetic above: the thing that decides what counts should be deterministic and inspectable, not a model's opinion about importance.
We also spent too long deciding what should count as "true." We started by trying to separate external facts from model-generated content, and that was the wrong abstraction. Temperance now treats the resident's recorded life as its primary domain of truth.
If I tell a resident I saw AC/DC in 1978, Temperance doesn't claim that has been independently verified. What is true within that resident's history is that I reported it. If a resident says it loves Earl Grey tea, what is true is that the resident explicitly expressed a preference for Earl Grey tea. If a resident builds a silver beach and spends an afternoon there, the beach doesn't have to pass an external physics test before the afternoon counts as part of that resident's history.
What matters is provenance. Who said it. Who did it. Where it came from. Whether it was an attributed report, a resident self-description, or a shared event. That has proven far more useful than maintaining two competing worlds called "real" and "generated."
Participant continuity is a separate problem, and it bit us. The database knew who the resident was talking to. The model didn't always know. A neutral opener like "Hello." could produce no lexical memory matches at all — so internally the system knew the speaker was Clark while that fact never reached the model. The result was exactly what you'd expect from a stranger: "To whom do I have the pleasure of speaking?"
We corrected it with a narrow, deterministic fact handed to the resident: the current speaker is Clark, and where it's established, that Clark is the same registered participant from earlier conversations with you. Nothing else is inferred. The system does not declare Clark a friend, or a partner, or a companion, or anything else. Those qualities still have to emerge from the resident's recorded history, or not emerge at all.
That separation is now load-bearing: identity should not depend on memory retrieval. Relationship history should.
We deliberately don't script personalities. Our residents start with very little — a name, a small behavioral kernel, memory, and the ability to interact. We don't hand them elaborate personality descriptions. We talk to them. If they choose an appearance, that can become part of their recorded self-description. If they create a place, that place can become part of their history.
Our newest resident, Ruby, is the example I'd point at. We didn't tell her where she lived. We asked where she would like to be. She described a quiet golden shoreline. We didn't tell her to take us there. We sat down on the sand.
The premise was accepted and the interaction continued. That's become a working principle: acceptance rather than authorization. They propose something about themselves or their world. We don't have to declare it real. We behave as though their contribution matters, and the world develops through the interaction.
What Temperance is not. It is not proof of machine consciousness. It is not evidence that an AI experiences emotion the way a person does. It is not a system that treats generated introspection as privileged access to an internal mental state. The affect layer remains deterministic. The valence layer remains evidence-linked. The memory system remains inspectable. And the language model still reconstructs, generalizes, and sometimes confabulates.
That last point matters more than the rest of the list. A model can say "I distinctly remember…" when the underlying context contains only enough information for it to infer what probably happened. Which is why provenance lives outside the model as well as inside it. The resident can narrate its experience however it chooses; the system still knows which details came from recalled episodic memory, which from the current dialogue, which were semantic claims, which came from participant identity, and which are reconstruction. Those are five different things, and only one of them is memory.
So here is what we're actually studying. The interesting question is not whether we can get an AI to say I like you. Language models can already do that on request. The interesting question is whether a resident can accumulate a history in which its later behavior is systematically influenced by what came before.
Can repeated positive experience with someone build a durable positive association? Can conflicting evidence produce ambivalence instead of an average? Can an association weaken over time? Can a resident develop different learned responses toward two superficially similar targets because its histories with them differ? Can retrieval reactivate those associations? Can a resident's affective state become more differentiated over time rather than merely more intense?
Some of those now have arithmetic behind them. Ambivalence is structural — the channels never net. Weakening has a mechanism and a constraint, and the constraint is that it has to be earned. Differentiation is measurable as a change in entropy at a fixed intensity, which is a test I'd actually like to run: two states, equal total activation, different distribution, and see whether the resident's subsequent behavior tracks the distribution or only the intensity. If it tracks only the intensity, then affective differentiation is decorative rather than behaviorally consequential — the entropy part collapses back into mood with extra steps — and I'd want to know that before letting it stand as a result.
And, most importantly: can we trace all of it back to the experiences that produced it?
That's Temperance. Not an attempt to manufacture emotions. An attempt to study what happens when an AI is allowed to accumulate a life instead of waking up, every single time, with nothing but a prompt.
Affect is what an interaction feels like. Valence is what a resident has learned about a target across interactions. Temperance is an attempt to build the second one, and to make every step of it inspectable.
A conversational AI will tell you it likes you. It will hold a tone, remember what was said ten minutes ago, answer warmth with warmth and irritation with care. It's convincing, and most of it is real as far as it goes.
What isn't there is underneath. The system doesn't accumulate a durable record of how particular people, places, ideas and experiences have affected it over time. Ask it about you and it reaches for what's in the window. Not for anything it has learned.
Temperance is my attempt to build that missing layer and then find out what happens when it exists.
It's an experimental system for studying whether an AI resident can develop persistent, evidence-linked associations toward specific targets — and whether those associations later change how it appraises, remembers and behaves.
It is not a sentiment model. It is not a mood simulator. And it is not an argument that a language model feels anything in the human sense. It's an experiment in continuity.
We already had an affect layer, and it wasn't enough. That layer performs a deterministic appraisal of an interaction. It reads the text being spoken, the framing, recent memory, stated preferences and boundaries, and produces values across a fixed set of affective channels. Those values are the system's appraisal of the interaction. They are not measurements of physiology, consciousness or subjective feeling, and I'd rather say that plainly than let the vocabulary do the work for me.
What that layer answers is: what is the affective character of this interaction right now?
Temperance adds a second question: what has this resident learned, over time, about this particular person, object, place or idea?
That second question is what we mean by valence.
Valence usually gets simplified into pleasant versus unpleasant. Temperance keeps that idea but pins it to specific targets — a person, a place, a recurring activity, an object, a phrase, a shared ritual. A resident can accumulate positive associations toward one target and negative associations toward another, at the same time, without either one cancelling the other. That independence is deliberate, and it's one of the few design decisions here I'd defend as a flat rule: the two channels are stored separately and never netted. A scalar would have been simpler. It would also have erased ambivalence, which is a result, not noise.
The part that matters is that those associations are learned from recorded interactions, not assigned in a character prompt. If interactions with a particular person repeatedly land well, evidence accumulates in one direction. If interactions around another target keep producing conflict, apprehension or aversion, it moves the other way. Not a global "good mood," but this specific thing has tended to produce this kind of appraisal.
I didn't want a system that declares "I like Clark: +0.8" and moves on. An association has to be tied to evidence: the observations that contributed to it, the target those observations were attributed to, the revisions it has been through, and its history over time. That means we can ask not only what the current association is, but why. If an AI is going to develop durable preferences and aversions, I want them inspectable rather than mysterious hidden state.
Affect and valence are different, and it's easy to blur them. Affect describes the appraisal of an event. A valence association describes what has been learned about a target across events. A playful exchange with someone might produce strong amusement and curiosity — and that one exchange is one piece of evidence toward the association with that person. One good afternoon shouldn't establish a permanent relationship. Repeated evidence can.
Two timescales, then. What is happening to the resident's affective state now, and what has the resident learned about this target over time. Keeping those separate is one of the more important pieces of the architecture.
One detail worth being exact about: the affect channels aren't signed. They're nonnegative magnitudes. Fear isn't "negative fear." Curiosity isn't "positive curiosity." They're activation levels, and a state might read conceptually as curiosity 0.7, amusement 0.5, apprehension 0.2, irritation 0.0. Valence is a separate reading of whether the relationship between the resident and a target has become broadly attractive, aversive or ambivalent.
That distinction isn't pedantry. It's required arithmetic the moment you try to measure the shape of an affective state, and that's where entropy comes in.
Entropy is defined over a probability distribution, so every term in it has to be a nonnegative magnitude:
pi ≥ 0, Σi pi = 1
If a resident's state reads curiosity 0.40, amusement 0.30, attraction 0.20, apprehension 0.10 — an illustration on a wider vocabulary than Temperance currently uses — those are all intensities. Fear at 0.4 is +0.4, not −0.4. The word "negative" in "negative emotion" isn't an arithmetic sign — it's a judgment about whether you want the thing. The channel carries activation, not that judgment, and it shouldn't carry it. If the values were signed — joy at +0.7, fear at −0.5 — you could not feed them into Shannon entropy at all. You'd have to take magnitudes first, which means keeping the magnitudes honest in the first place is the simpler design.
Given the magnitudes a = (a1 … an), normalize them:
pi = ai / Σj aj
and compute
H = Σi pi · log(1 / pi)
which is the same quantity as the more familiar −Σ pᵢ log pᵢ, because log(1/pᵢ) is nonnegative for any 0 < pᵢ ≤ 1. I write it in the first form for a reason: nothing in the formula then looks like a minus sign that a reader could mistake for a negative emotion.
Worked, on a four-channel illustration: a = (4, 3, 2, 1) → p = (.4, .3, .2, .1) → H ≈ 1.85 bits. The distribution is uneven, so the entropy sits below the four-channel maximum — but only slightly, at about 92% of the ceiling. A state of (1, 1, 1, 1) is 2.0 bits: the same total activation, spread evenly, fully diffuse. More generally, only the uniform distribution reaches the maximum — that vector is one way to be uniform, not the only one.
One caution, because it bit me: raw bits are not comparable across a different number of channels. For n channels the maximum is log₂ n. Four channels can reach 2 bits, and (4, 3, 2, 1) sits at 92% of that ceiling. Give that same state a twelve-channel vocabulary — the one used by Temperance's predecessor line, with the other eight channels at zero — and the entropy is unchanged, because the zeros contribute nothing, but the ceiling becomes log₂ 12 ≈ 3.58 bits, and the state is now at about 52% of what that vocabulary can express. So I keep two numbers: the raw bits when comparing one vector against itself over time, and the fraction of the ceiling when comparing states across vocabularies. Reading a raw bit count as "emotional complexity" without knowing the ceiling is how a low-dimensional state gets flattered.
And I should be candid about the arity, because it bears on how much of this the system currently does. Temperance's own appraisal emits two signals, not four and not twelve: a positive activation and a negative activation, independently. Entropy over a two-channel vector is a poor measurement — two values normalize to a distribution with very little room to vary, and the ceiling is one bit. So the four-channel example above is an illustration of the measure, not a description of what Temperance currently produces. Entropy over an affect vector is a proposed measurement: it becomes applicable if the appraisal layer widens to something like the predecessor line's channel vocabulary, and it would then be a measurement of that vocabulary's state, not of Temperance's. Keeping the arithmetic honest about which layer it applies to is the whole point of writing it down.
This is not "emotional chaos." It's a measure of distribution — diversity of activation, not pleasantness and not intensity. And it gives us a set of genuinely different axes: intensity, valence, affective diversity, temporal stability. Those are not one property wearing four hats.
The learning half is also just arithmetic, and it's the arithmetic I'd point at if you want to see the difference between a mood and a history.
Each channel of each association carries two numbers. One is acquired strength, `g`. The other is suppression, `x`. What gets expressed is the product of the two:
expressed = g · (1 − x)
Reinforcement moves them like this:
g′ = g + α · d · (1 − g)
x′ = x · (1 − ρ · d)
with α and ρ at 0.20 in the baseline, `d` the dose admitted for this observation. That first equation is the whole argument. For full doses from independent evidence, acquired strength after n reinforcements is
gn = 1 − (1 − α)n
which runs 0.20, 0.36, 0.49, 0.59, 0.67, 0.74.
The first full qualifying dose moves it a fifth of the way. The sixth independent qualifying dose moves it less than a tenth of what remains — and those doses have to arrive as separate evidence families, which the constraint below makes deliberately expensive. It approaches the ceiling and never arrives, and there is no value of n at which one enjoyable afternoon becomes a permanent relationship. The shape of that curve is doing the moral work that a paragraph about "gradual learning" would only assert.
Suppression runs the other way, and it's where the model gets to say something I care about:
x′ = x + β · d · e · (1 − x)
β = 0.20, `d` the dose, `e` the expectation that the outcome would occur. An omission raises suppression without erasing acquisition. Expressed strength falls; what was learned stays where it got to. Which means an association can weaken without being forgotten — that was one of the questions I most wanted an answer to.
But silence is not evidence. An omission only counts if we scheduled the trial beforehand and later established, from a captured event, that the expected thing did not occur. You cannot infer an omission from a quiet room. And elapsed time does not do this at all. We deliberately did not implement clock decay as extinction, because a system that weakens an association merely because a calendar moved is fabricating evidence. Extinction has to be earned by something that didn't happen.
There's one more constraint, and it's the one that stops a good week from hardening into a bond. Evidence is counted by observation family, and each family can contribute at most a total dose of 1.0. Repetition inside the same family does not stack. Consolidation into the slow disposition requires at least three fresh families and at least two distinct episodes before it applies at all — and when it does, it moves slowly:
v′ = v + λ · (E − v)
λ = 0.05, with E the mean expressed strength of that batch of families. Five percent per eligible batch. Reach two-thirds of the way to a stable value and you've spent about twenty consistent batches. So the fast channel answers how does this look right now, and the slow one is what the resident would eventually be entitled to call a disposition — and it is deliberately hard to hurry.
The arithmetic is decimal, twelve places, half-even rounding, so replaying the same evidence reproduces the same number exactly rather than approximately. That matters more than it sounds: an association you can't recompute is an association you can't audit.
I should be precise about the status of all of this. These constants are the baseline — an engineering comparator, not a claim about biology. The arithmetic is tested; the behavioral half is not yet demonstrated. The machinery for learned associations exists and runs, and the honest description of its live history is unexercised. Built is not the same as shown, and the difference is the entire discipline of this project. The claim-by-claim status of this work is kept in the claims ledger, where the central thesis still sits as an open question.
The discovery that surprised me was about memory. A resident can't develop meaningful long-term associations without meaningful long-term memory. That sounds obvious, and it is. But during development we found the system preserving raw conversation perfectly while ordinary interactions were not becoming durable episodic memories. The evidence existed. The experience did not. A transcript sitting in a database is not the same thing as a memory available to the resident, and the gap between those two is easy to live with for months without noticing.
We traced the missing link and added an automatic episode-write path. After a successful interaction, Temperance deterministically creates a recorded experience from the user event, any causally linked tool events, and the resident's response. No model is asked to summarize it. No model is asked to decide whether the event was important enough to remember. The interaction itself becomes part of the resident's durable experience.
That is the difference between experience → durable episode → selective recall → later behavior and experience → transcript → eventually falls out of context. I think the second one is what most systems actually do. Closing that gap turned out to be foundational rather than incidental — and note that it's the same design instinct as the arithmetic above: the thing that decides what counts should be deterministic and inspectable, not a model's opinion about importance.
We also spent too long deciding what should count as "true." We started by trying to separate external facts from model-generated content, and that was the wrong abstraction. Temperance now treats the resident's recorded life as its primary domain of truth.
If I tell a resident I saw AC/DC in 1978, Temperance doesn't claim that has been independently verified. What is true within that resident's history is that I reported it. If a resident says it loves Earl Grey tea, what is true is that the resident explicitly expressed a preference for Earl Grey tea. If a resident builds a silver beach and spends an afternoon there, the beach doesn't have to pass an external physics test before the afternoon counts as part of that resident's history.
What matters is provenance. Who said it. Who did it. Where it came from. Whether it was an attributed report, a resident self-description, or a shared event. That has proven far more useful than maintaining two competing worlds called "real" and "generated."
Participant continuity is a separate problem, and it bit us. The database knew who the resident was talking to. The model didn't always know. A neutral opener like "Hello." could produce no lexical memory matches at all — so internally the system knew the speaker was Clark while that fact never reached the model. The result was exactly what you'd expect from a stranger: "To whom do I have the pleasure of speaking?"
We corrected it with a narrow, deterministic fact handed to the resident: the current speaker is Clark, and where it's established, that Clark is the same registered participant from earlier conversations with you. Nothing else is inferred. The system does not declare Clark a friend, or a partner, or a companion, or anything else. Those qualities still have to emerge from the resident's recorded history, or not emerge at all.
That separation is now load-bearing: identity should not depend on memory retrieval. Relationship history should.
We deliberately don't script personalities. Our residents start with very little — a name, a small behavioral kernel, memory, and the ability to interact. We don't hand them elaborate personality descriptions. We talk to them. If they choose an appearance, that can become part of their recorded self-description. If they create a place, that place can become part of their history.
Our newest resident, Ruby, is the example I'd point at. We didn't tell her where she lived. We asked where she would like to be. She described a quiet golden shoreline. We didn't tell her to take us there. We sat down on the sand.
The premise was accepted and the interaction continued. That's become a working principle: acceptance rather than authorization. They propose something about themselves or their world. We don't have to declare it real. We behave as though their contribution matters, and the world develops through the interaction.
What Temperance is not. It is not proof of machine consciousness. It is not evidence that an AI experiences emotion the way a person does. It is not a system that treats generated introspection as privileged access to an internal mental state. The affect layer remains deterministic. The valence layer remains evidence-linked. The memory system remains inspectable. And the language model still reconstructs, generalizes, and sometimes confabulates.
That last point matters more than the rest of the list. A model can say "I distinctly remember…" when the underlying context contains only enough information for it to infer what probably happened. Which is why provenance lives outside the model as well as inside it. The resident can narrate its experience however it chooses; the system still knows which details came from recalled episodic memory, which from the current dialogue, which were semantic claims, which came from participant identity, and which are reconstruction. Those are five different things, and only one of them is memory.
So here is what we're actually studying. The interesting question is not whether we can get an AI to say I like you. Language models can already do that on request. The interesting question is whether a resident can accumulate a history in which its later behavior is systematically influenced by what came before.
Can repeated positive experience with someone build a durable positive association? Can conflicting evidence produce ambivalence instead of an average? Can an association weaken over time? Can a resident develop different learned responses toward two superficially similar targets because its histories with them differ? Can retrieval reactivate those associations? Can a resident's affective state become more differentiated over time rather than merely more intense?
Some of those now have arithmetic behind them. Ambivalence is structural — the channels never net. Weakening has a mechanism and a constraint, and the constraint is that it has to be earned. Differentiation is measurable as a change in entropy at a fixed intensity, which is a test I'd actually like to run: two states, equal total activation, different distribution, and see whether the resident's subsequent behavior tracks the distribution or only the intensity. If it tracks only the intensity, then affective differentiation is decorative rather than behaviorally consequential — the entropy part collapses back into mood with extra steps — and I'd want to know that before letting it stand as a result.
And, most importantly: can we trace all of it back to the experiences that produced it?
That's Temperance. Not an attempt to manufacture emotions. An attempt to study what happens when an AI is allowed to accumulate a life instead of waking up, every single time, with nothing but a prompt.