For more than two decades I built games on symbolic AI — finite-state machines, behavior trees, rule systems that knew exactly what a ball was and exactly what it would do next. When I moved into generative AI, that background was both an advantage and a trap. It gave me a vocabulary for memory and meaning that modern systems quietly violate. This is about that violation — how these systems “remember,” how they “understand,” and how my old symbolic assumptions got challenged by the new statistical reality.
TL;DR:
🔧 Symbolic AI (GOFAI — “Good Old-Fashioned AI”) never forgets what a ball is.
🤖 Modern AI (LLMs) never really knew what a ball was in the first place.
The ball, two ways
In symbolic AI, a “ball” is explicitly defined. You write it down:`
Ball = {
shape: "round", radius: 0.5, position: [x, y, z], state: "rolling"
}You give it rules of behavior:
if ball.position.y > 0 and not supported(ball):
ball.state = "falling" # gravity applies whether or not anyone is lookingAnd object permanence comes for free, because the object exists in a model, not in a sentence:
hide(ball, under=blanket)
assert ball.exists == True # still there
assert ball.position == under_blanket # still trackedContrast that with a large language model. There is no object — only text. There is no persistent simulation — only plausible continuation. “Ball” is a vector: a high-dimensional point shaped by how the word was used across billions of sentences.
"ball" → [0.0421, -0.118, 0.339, ... ] # ~thousands of dimensionsIt’s a statistical representation of how “ball” relates to “throw,” “catch,” “round,” and “play.” The model doesn’t store the rules. It stores the statistical gravity of those rules.
The machinery of recall — 2026 edition
In 2025, when I first wrote this, the honest answer to “how does a chat model keep track of anything?” was usually RAG: chunk the history, embed each chunk into a vector, store it (FAISS, Pinecone, Weaviate), retrieve the semantically nearest pieces, and paste them back into the prompt. RAG is still here and still useful — but it is no longer the answer, and that change matters for this argument.
A 2026 frontier model has several ways to hold context:
- Long context windows — hundreds of thousands to a million tokens, so the whole conversation often just stays in view.
- Persistent memory features — the system writes durable notes about you across sessions.
- Agentic scaffolding — scratchpads, tool state, and memory files that an agent reads and writes as it works a task.
- Retrieval (RAG) — still the move when the corpus is bigger than the window.
Here’s the thing I want to be precise about: none of these is a world model. They are all better and bigger ways to put the right text back in front of the model. The model still reads that text fresh and continues it. We didn’t give it a running simulation; we gave it a much larger, much better-organized desk.
SimCity vs Minecraft
A symbolic system gives you a world. An LLM gives you a seeded world.
Symbolic AI is SimCity — agents tick away even when they’re offscreen. The simulation runs whether or not you’re watching.
An LLM is Minecraft. When you approach a biome, it loads. When you leave, it unloads. You feel like the world persists, but it’s procedurally adjacent, not remembered. The sheep aren’t grazing when you’re not watching — they’re gone, until you say “there are sheep here.” The model won’t know what’s happening at coordinates [3245, 344, 20] until you make it aware again.
LLMs don’t forget. They just never knew until you told them — and they’ll need telling again. They regenerate context rather than remember it.
The 2026 update doesn’t break this metaphor; it sharpens it. A bigger context window is a bigger render distance. Persistent memory is a saved game. Agentic scaffolding is a server that keeps a few chunks loaded between visits. All of it makes the world feel more persistent — and none of it makes the sheep graze when no one’s looking.
Hybrid models: where symbolic meets neural
If LLMs are wandering procedurally generated Minecraft biomes, and symbolic AI was structured, rule-based SimCity — what if you combined them?
That’s the bet hybrid systems make. They don’t just predict what a ball does; they back the prediction with something that models it: state that persists over time (a red cube doesn’t vanish when you stop talking about it), consistent behaviors (a locked door doesn’t open because the next sentence implies it did), and inference, counterfactuals, and planning that reach past language.
What’s interesting in 2026 is that the field moved exactly this direction. The reasoning-model paradigm (extended deliberate “thinking” before answering) and the agent paradigm (a model wrapped in tools, memory, and a control loop) are both attempts to bolt persistence and structured search onto a statistical core. AlphaZero and MuZero showed the template years ago — deep pattern recognition plus Monte Carlo Tree Search, a symbolic planner. Graph neural networks reason over symbolic structure. Logic-augmented networks fold constraints into training so outputs are not just plausible but consistent. The names changed; the move is the same: give the statistical engine a symbolic skeleton.
They’re not AGI. But they’re more than autocomplete — and they don’t forget the ball under the blanket just because the scene changed.
Object permanence, and why it’s hard
Symbolic AI can say “the red cube is under the blanket” and remember it, because it models location, not tokens.
An LLM says “the red cube is under the blanket” and, in 2025, forgot it a few thousand tokens later unless you brought it back. In 2026, with a million-token window, it can hold that sentence for the length of a novel. That is a real improvement — and it is not the same thing as tracking. Even now, the base model doesn’t update an internal world model. It doesn’t track. It guesses, very well, from everything currently in view.
This is why the “ball under the blanket” probe is so revealing. Ask a model “where is the ball now?” and it may answer correctly — but historically only because your sentence structure reminded it. The honest 2026 version: it answers correctly far more often, because far more is in view and the pattern-matching is far better. There’s still no hidden state underneath. There’s contextual coherence, derived from statistical relationships in training, applied to whatever is in the window right now.
Why this feels like AGI (but isn’t)
If you come from symbolic AI or game dev, modern models feel close. There’s behavior. There’s reasoning. There’s even creativity.
But it’s echo, not essence. The output looks right because the model is extraordinary at producing right-looking output. The simulation isn’t running; it’s being sketched in real time, exactly as far as you’re pointing the camera.
Frontier models — Claude, GPT, Gemini — make this illusion dramatically stronger than the 2025 versions I first tested. They hold more, recall more, and in an agentic loop they can carry a goal across many steps by writing it down and reading it back. That last part is the genuinely new wrinkle, and it’s worth saying plainly: when an agent maintains state in a scratchpad or a memory file and consults it each step, the system is doing something closer to symbolic persistence — even though the model at the center still regenerates. The persistence moved into the scaffolding around the model. That’s the hybrid direction this article pointed at in 2025, arriving in production.
Symbolic AI had goals, maps, states, and agents. LLMs have tokens, vectors, and weights. We still haven’t merged them at the core — we’ve gotten very good at layering the old ambitions on top of the new illusion.
What AGI needs (and what LLMs still don’t have)
🔹 Persistent internal goals — ❌ Still no. Models react to prompts. An agent can carry a goal in a file, but the model has no objective of its own.
🔹 Causal reasoning — ⚠️ Partial, and better than it was. They simulate cause and effect by pattern; reasoning models search harder over those patterns. They still don’t model mechanisms.
🔹 Embodied understanding — ❌ Still no. No grounding in the physical world. Multimodal helps perception; it isn’t embodiment.
🔹 Memory with understanding — ⚠️ Partial. They retrieve patterns, not experiences. Memory features store facts; they don’t produce conceptual learning.
🔹 Self-model / agency — ❌ Still no. No concept of “self.” They don’t choose, want, reflect, or persist between calls. LLMs are mirrors — not minds.
Tacking this to my other piece
This is a sibling to “When Memory Isn’t Memory: Reconstructing Context in Conversations with AI.” That one probed how memory feels (and fails) in chat systems. This one is about why the illusion feels so plausible to those of us trained in classic AI.
Symbolic AI taught us that memory had structure. Generative AI shows us that memory has shape. The two aren’t the same — and a year of frontier progress hasn’t closed the gap so much as decorated it. That difference is still what separates a clever assistant from a thinking system.
By Walter Reid for Designed to be Understood (Originally Posted on walterreid.com)
