The Rival Theory

Second Opinion

Everything looks inevitable until somebody does it differently.

In 2021, John Laird - who built Soar - published a long, careful, admirably fair comparison of his architecture against ACT-R. Read it knowing he invented one of the contestants.

It is the most useful document in this literature. It keeps pointing at things ACT-R does that ACT-R never bothered to justify, because from the inside they never looked like choices.

Where they agree: the shape of a mind

Picture a small office with one clerk in it. Behind one door is an archive of facts. Behind another is a workshop full of habits - the if-this-then-that rules from part one. In the middle is a single desk, which holds whatever the clerk is working on right now, with an in-tray for whatever the senses deliver and an out-tray for whatever the hands are told to do.

Both theories draw the same office: memory for facts, memory for skills, a small workspace (the buffer from part one), and the senses on either end.

    skills  ←──────────────────→  facts
        │                           │
        └──────  what you are  ─────┘
                thinking about now
                 │            │
              senses       actions

The office has one clerk, and the clerk has two hands. Both theories insist on that bottleneck: one memory retrieval at a time, one rule firing at a time. They put it there because the evidence about people supports it.

Try it now

Recall the name of your first teacher and your childhood phone number, and try to have both arrive at the same instant rather than one after the other.

You will find they queue. One arrives, then the clerk goes back for the other. That is quietly damning for any AI system that fires off twenty parallel searches per question and calls the result thinking.

No Own Goals

You are not allowed to see your own scores.

Laird draws a line nobody else in this field bothers with. There is what you know, and there is the bookkeeping about what you know - how often you have used it, how recently, how reliable it has turned out to be.

The bookmaker from part two keeps an odds board on every fact, pricing how likely you are to need it tomorrow. The odds board is the bookkeeping, and activation is what is chalked on it. In both theories, the board sits behind a wall.

You cannot see your own activation numbers, and you cannot set them. You can only behave - and the bookkeeping updates itself behind your back.

Try it. Pick a word you have no reason to need next week - marmalade will do - and decide, as firmly as you can, that it matters. Nothing on the board moves. Now say it out loud five times. That moves it, because saying it is using it, and every use rings its bell.

That limitation is what gives the number its meaning. A credit history means something because you cannot simply announce that you have excellent credit.

A large and growing number of AI memory systems hand the model direct write access to its own importance scores, which looks like learning. Laird's objection is mechanical:

Looking is probably fine

Showing something its own confidence level is defensible - a punter who gets a glance at the odds board. It would have to learn what the number means, but nothing underneath breaks.

Touching breaks everything

Let something edit its own use-counter - ring its own bells whenever it likes - and the counter stops meaning "how often this was genuinely useful", which is the only thing the entire equation assumes about it.

The rule is look, never touch, and both theories arrived at it independently.

Where They Split

Soar lets a thought travel further than ACT-R does.

Back to the crowded room from part three. You shout a name, the person it belongs to turns round, and that is a memory lending strength to its neighbour. The two theories split over what happens next: whether Priya, having turned, nudges the person beside her, who nudges the next, like a whisper passed down a row.

In ACT-R, nobody passes it on: spreading goes strictly one step. Whatever is on your mind lights up its immediate connections and stops dead.

In Soar, the whisper travels, for as many steps as you configure, getting fainter at every hop - each hop being one handoff from one memory to the next.

So any system that walks a knowledge graph three hops out, fading as it goes, has built something closer to Soar - whatever its citations claim. That is not wrong: multi-step is very likely what you want for a web of documents. But it means the per-hop fade - how much quieter the whisper gets at each handoff - has no established value to borrow, because the theory everybody cites does not contain the mechanism at all.

We spent months doing Soar and calling it ACT-R.

Soar remembers days. ACT-R only remembers facts.

There is a difference between knowing that Paris is the capital of France and remembering the afternoon you learned it. The first is a fact, an index card in the archive. The second is an episode - a snapshot of everything that was happening at once. Psychologists call the two semantic and episodic memory.

Try it: think of the last meal you ate out. What comes back is a scene - the table, who was there, roughly what was said. Now step forward to what you did after you left. You can, a little shakily, because an episode comes with a before and an after attached.

Soar keeps both, as separate systems. ACT-R has only the first. You can hand-encode time into facts, but it costs effort every time and nothing captures the whole scene automatically.

Soar's episode store can then do things a pile of facts simply cannot: find the most recent time that matched, fall back to the closest match when nothing fits exactly, and step forwards and backwards through time, so an experience can be replayed in order.

"What else was going on around the time we decided that?" is a question a ranked list of facts answers badly and a timeline answers instantly. Choose ACT-R and you inherit the blind spot.

Where They Converged

Both theories ended up building the same guardrails.

Two teams, with different goals, spent forty years arriving at the same place independently.

Suppress whatever you just used

In both systems, a memory you just retrieved gets temporarily held down, specifically so you do not keep pulling up the same thing over and over.

ACT-R's version has the best name in the field: declarative finsts, short for "fingers of instantiation." You have a handful of fingers. You keep one on each of the last few things you grabbed. Run out of fingers and the oldest one lifts off.

You can feel the fingers working. Name five fruits, quickly. Apple is almost certainly the most practised fruit you own, and it still does not come back a second time, because there is a finger on it.

If using a memory strengthens it, and nothing holds down the ones you just used, then:

Every time a memory surfaces it gets stronger, and being stronger makes it the first thing the next question grabs.

The winners keep winning, and the gap widens every turn.

Every individual answer is defensible, so nothing looks broken while the system slowly collapses into repeating itself. The only way to notice is to go looking for it.

Never delete anything

Laird, on both systems: there are no forgetting mechanisms for facts. Things can sink so deep that nothing ever retrieves them.

Forgetting is loss of access with the content left intact. In the swimming pool from part one, the object is still down there in the dark.

Which means every recent AI memory paper announcing "make it inaccessible, don't delete it" as its headline insight is rediscovering a position these two settled in about 1983.

Soar does forget in exactly one place, and it runs the opposite way to the bike in part one: learned skills decay, facts never do. For machine memory the asymmetry is worth keeping. A knack you stop using can fade away entirely. A thing you know just sinks.

Do not trust anything new

Picture the workshop of habits taking on a new hire. On day one the new hire cannot overrule the foreman who has run the place for twenty years, however confident the new hire sounds. Both theories ended up running a probation period.

A freshly learned rule in ACT-R starts deliberately weak, so it cannot immediately elbow aside the thing it was derived from. Soar originally let new rules compete straight away, and had to be changed so that rules must be learned several times over before they take effect.

So if you are building something where a brand-new memory arrives at full strength and can immediately outrank things you have known for years - two independent teams already ran that experiment, and both came back and changed it.

A Third Opinion

The shape of your theory decides what you can think about.

A third architecture, CLARION, makes one uncomfortable observation. Picture two buildings. In one, the ground floor is raw pattern-matching - "this feels like that" - and the tidy symbols live upstairs. In the other, the symbols are the building and pattern-matching is a conservatory somebody added later.

CLARION is the first building. ACT-R is the second: similarity is an optional bolt-on. A review comparing the two in practice found this:

The structure of the architecture mostly determines which ideas anyone bothers to express in it.

The other ideas remain possible but awkward, and that is enough to keep them out.

Choosing ACT-R pushes you towards explaining relevance in terms of frequency, recency and connection, and away from explaining it in terms of raw similarity. CLARION's answer to anybody bolting a similarity score onto a history score - the metres-and-kilograms problem from part two - would be: stop doing that and put similarity at the bottom where it belongs. Knowing that alternative exists is what stops "this bit was awkward" from being mistaken for "this bit is just hard."

Baggage Allowance

Only the ranking equations carry over to machine memory.

What crosses over cleanly

The memory equations, almost perfectly. They are a theory about ranking, and everything they need - when things were used, how often, what connects to what - is already sitting in a column in somebody's database. The bookmaker only ever needed the betting slips, and databases are full of them.

Two ideas cross over as well: that failing to remember should be a visible, first-class outcome, and the split between knowing things and knowing how to do things.

What does not

Everything about how knowledge is actually stored. ACT-R's memories are neat inspectable objects. A language model's knowledge is smeared across billions of numbers with no edges. Bolting a database onto one gives it a prosthetic memory.

Consolidation and reasoning do not cross over either, except as metaphor.

We spent the last few years building machines that can hold a conversation, and gave them memory systems in which everything is equally available forever, nothing is genuinely forgotten, and there is no way to say "I don't know." That is a swimming pool one inch deep, with everything floating at the surface at once.

Meanwhile the actual problem had been studied to death - and fitted to three decimal places against stopwatches. Two rival theories spent forty years arguing each other into a set of guardrails that neither of them started with.

Every one of those guardrails is a mistake somebody else already made, wrote down, and fixed.

Part five covers building the scorer and measuring it offline: every variable, what it does, what getting it wrong cost us, and what comes next. One constant turned out to disable the history mechanism entirely, and the test corpus had absorbed its own answer key.

Activation in Practice →