INQUIRING LINE

Inside a neural network, memorizing facts and reasoning through problems are completely separate jobs — learned from different data.

Why does knowledge storage separate from reasoning circuits in neural networks?

This explores why, inside a neural network, the machinery that recalls facts appears to be physically and functionally distinct from the machinery that reasons — and what that separation tells us about how these models actually work.


This explores why the parts of a network that store facts seem to be distinct from the parts that reason, and the corpus offers a surprisingly consistent picture: the two rely on different kinds of learned structure, so they end up living in different places. The clearest evidence comes from tracing what pretraining documents actually drive each behavior. When researchers looked across five million pretraining documents, factual recall turned out to depend on narrow, document-specific memorization — the model essentially points back to the one place it saw the fact — while reasoning drew on broad, transferable *procedural* knowledge spread across many diverse sources Does procedural knowledge drive reasoning more than factual retrieval?. Storage is lookup; reasoning is a reusable routine. Those are different jobs, learned from different data signatures, so it makes sense they don't share the same circuitry.

You can watch the seam directly in how models handle logic. Syllogistic reasoning runs through a content-independent three-stage circuit — recite, suppress the middle term, mediate — that works the same across architectures. But separate attention heads carrying world knowledge sit alongside it and bias the conclusion toward what's *plausible* rather than what's logically valid How do language models perform syllogistic reasoning internally?. That's the separation made visible: a reasoning engine and a knowledge store operating as distinct components, sometimes even fighting each other. Notably, the knowledge contamination gets *worse* at larger scale, which suggests the two systems don't merge as models grow — they stay separate and the stored-knowledge pull gets stronger.

There's a deeper reason this modularity shows up at all. Networks appear to naturally decompose compositional tasks into isolated subnetworks — pruning experiments show you can ablate one subroutine and only its function breaks, and pretraining sharpens this modular structure Do neural networks naturally learn modular compositional structure?. So separation isn't something engineers impose; it's a structure gradient descent tends to find on its own. It also explains why a 7-million-parameter network that just recurses on its own latent reasoning state can out-generalize billion-parameter LLMs on hard puzzles — the reasoning gains come from the recursive *process*, not from a giant pile of stored knowledge Can tiny recursive networks outperform massive language models?. Reasoning and storage scale differently because they're doing different things.

The most useful reframing in the corpus maps this onto the brain. Transformer weights behave like a distributed neocortex holding consolidated, slow-to-change knowledge; retrieval systems act like the hippocampus doing fast new encoding; and agentic state plays the prefrontal role of executive control Can brain memory systems explain how LLMs should store knowledge?. Under that lens the separation isn't a quirk — it's the same division of labor evolution landed on, because consolidated knowledge and active manipulation have opposing requirements (stable vs. fast, general vs. situation-specific).

Here's the twist worth leaving with: the separation is real but leaky, and the leak is a liability. When semantic content is stripped from a reasoning task, LLM performance collapses even when the correct rules are handed to them in context — meaning much of what looks like clean reasoning is actually stored semantic association doing the work Do large language models reason symbolically or semantically?. So knowledge and reasoning are separable enough to have distinct circuits, but entangled enough that the model leans on memorized meaning whenever the pure-reasoning path gets hard. The open frontier — which the brain-memory framing names explicitly — is the *consolidation* mechanism that would let the two integrate cleanly instead of interfering.


Sources 6 notes

Does procedural knowledge drive reasoning more than factual retrieval?

Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.

How do language models perform syllogistic reasoning internally?

LLMs implement a content-independent three-stage reasoning mechanism—recitation, middle-term suppression, mediation—that works across architectures. However, additional attention heads encoding world knowledge systematically bias conclusions toward semantically plausible rather than logically valid answers, with contamination increasing at larger scales.

Do neural networks naturally learn modular compositional structure?

Pruning experiments reveal that neural networks implement compositional subroutines in isolated subnetworks, with ablations affecting only their corresponding function. Pretraining substantially increases the consistency and reliability of this modular structure across architectures and domains.

Can tiny recursive networks outperform massive language models?

A 7M-parameter two-layer network recursing on its latent reasoning state reached 45% on ARC-AGI-1, beating larger LLMs with 0.01% of their parameters. The gains come from recursion itself, not scale or hierarchical architecture.

Can brain memory systems explain how LLMs should store knowledge?

Research shows transformer weights function as a distributed neocortex for consolidated knowledge, RAG stores as hippocampal indexing for rapid encoding, and agentic state as prefrontal executive control. The CLS framework predicts why hybrid systems outperform single-tier approaches and identifies missing consolidation mechanisms that prevent memory integration.

Show all 6 sources
Do large language models reason symbolically or semantically?

When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are an interpretability analyst investigating a still-open question: why does knowledge storage separate from reasoning circuits in neural networks — and how might the two integrate cleanly?

What a curated library found — and when (dated claims, not current truth; findings span ~2023–2026):
- Across 5M pretraining documents, factual recall traces to narrow, document-specific memorization while reasoning draws on broad, transferable procedural knowledge — different data signatures, different circuits (~2024).
- Syllogistic reasoning runs a content-independent three-stage circuit (recite → suppress middle term → mediate), but separate world-knowledge heads bias conclusions toward plausibility over validity, and this contamination worsens at larger scale (~2024).
- Networks self-decompose compositional tasks into modular subnetworks; ablating one subroutine breaks only its function, and pretraining sharpens this (~2023).
- A 7M-parameter net recursing on latent reasoning state out-generalizes billion-parameter LLMs on hard puzzles — gains from process, not stored knowledge (~2025).
- Strip semantics from a reasoning task and performance collapses even with correct rules in context — memorized meaning is doing the work (~2023).

Anchor papers (verify; mind their dates): Structural Compositionality (arXiv:2301.10884, 2023); In-Context Semantic Reasoners (arXiv:2305.14825, 2023); Reasoning Circuits / Syllogistic (arXiv:2408.08590, 2024); Procedural Knowledge in Pretraining (arXiv:2411.12580, 2024); Recursive Reasoning with Tiny Networks (arXiv:2510.04871, 2025).

Your task: (1) Re-test each constraint — for every finding, judge whether newer models, training, tooling, orchestration (memory, retrieval, multi-agent), or evaluation has RELAXED or OVERTURNED it; separate the durable question from the perishable limitation, cite what resolved it, and say where a constraint still holds. (2) Since this is synthesis-probing, reconcile the threads — does the neocortex/hippocampus/prefrontal framing (arXiv:2601.09113) actually unify them, and what consolidation mechanism would let storage and reasoning integrate instead of interfere? (3) Propose 2 research questions assuming the regime has moved. Cite arXiv IDs; flag anything you cannot ground in a real paper.