A long-standing idea. A new research program.

Intelligence-as-search is not new. Allen Newell and Herbert A. Simon made heuristic search central to intelligent problem solving; Stuart Russell and Eric Wefald formalized choosing among computations; Richard Sutton identified search and learning as general methods that scale; and AlphaZero combined learned guidance with live search. Cognitive science has likewise studied interactions among habit and planning, episodic and gradual learning, and cognitive control. Agent architectures such as CoALA now bring memory, reasoning, and action into a shared decision process.

Our position is architectural, not a priority claim: language-model agents make this integration experimentally tractable. How should lessons in weights, particular experiences in memory, and new search in the moment form a coherent, compounding system?

We use cognition here in a functional, architectural sense: the organization that makes learning, memory, evaluation, and deliberation coherent. In that sense, cognition is coordination.

Here, search means resolving among possible actions, explanations, or internal trajectories—not exhaustively enumerating them. Brute force is only the limiting case when useful guidance is weak.

Intelligenceis search.

Reinforcement learninggeneralizes useful search into weights.

Episodic memorypreserves particular experiences and outcomes.

Explicit searchresolves what remains unknown.

Cognitioncoordinates them across situations and timescales.

Part I

Search Across Time

Figure 01 · Three regimes

The work of search can be distributed and reused across time.

Explicit search constructs and evaluates candidates at decision time. Reinforcement-learning post-training explores and evaluates trajectories, then amortizes useful regularities into model weights. Episodic memory retrieves an addressable record of a particular prior experience.

  1. 01Explicit search

    At decision time

    Construct and evaluate candidates for the current problem. Brute force is the limiting case.

    Search is paid for
    At the moment of decision
    Outcome carried forward
    A solution for the current case
    Example · Finding your keysInspect each plausible location
  2. 02Reinforcement learning

    Amortized into weights

    Explore and evaluate during reinforcement-learning post-training; amortize useful regularities into model weights.

    Search is paid for
    During RL post-training
    Outcome carried forward
    Generalized priors in weights
    Example · Finding your keysCheck where they are usually found
  3. 03Episodic memory

    Retrieved from an episode

    Retrieve details from a particular prior situation, trajectory, or outcome.

    Search is paid for
    During a prior experience
    Outcome carried forward
    An addressable episode
    Example · Finding your keysRecall the couch and table, then verify
CognitionCoordinates when and how each regime contributes.
These regimes are complementary. Cognition coordinates when and how each contributes to the same problem.

Explicit search at decision time

The first regime constructs and evaluates candidates while the current problem is being solved. Imagine that your usual route home is blocked. Neither habit nor a particular memory contains the answer to this exact configuration of roads and traffic, so you must consider the routes available now and construct a path you may never have taken before.

This is the strength of explicit search: it can construct solutions for novel configurations. Its weakness is cost. Brute force sits at the extreme—enumerate and evaluate everything reachable—but useful knowledge can structure the space and focus the search.

Amortize search into weights

The second regime generalizes useful regularities discovered through exploration and evaluation. During reinforcement-learning post-training, a system generates trajectories, receives feedback, and updates its policy. Patterns associated with better outcomes become more likely, allowing the model to produce a promising action at inference without repeating every exploration that shaped it.

In this sense, reinforcement learning is amortized search: some of the cost is paid earlier and its benefit is spread across future decisions. The result is a form of parametric memory, but not usually an addressable record of the trajectories themselves. Useful regularities are distributed through the model's weights.

Preserve a particular experience

The third regime is episodic memory: an addressable trace of a particular event, trajectory, or evaluated outcome. A learned prior might direct you toward where your keys are usually found; an episode from this morning might instead retrieve the couch and dining table. A precise memory may supply a single candidate that only needs to be verified.

RL-trained weights preserve regularities across many experiences; episodic memory preserves particular experiences in an addressable form. Both can reduce the need to search from scratch, but they differ in what they retain and how it is accessed.

Part II

AI Has Repeatedly Redistributed Search

Figure 02 · A timeline of where search happens

Each system divides the work differently between live search and learned guidance.

01
Explicit search dominant

Deep Blue

Deep Blue operated near the explicit-search end of the spectrum. Engineered evaluation, selective search, and expert game databases guided enormous computation, but much of the work still happened while choosing the next move.

02
–Learned memory + explicit search

AlphaGo → AlphaZero

A policy proposed promising moves. A value function evaluated positions. Tree search explored what the networks had not resolved. In AlphaZero, search produced better learning targets; the updated network then guided search for the next move.

Useful regularities from search were amortized into weights. Those weights made search for the next move more efficient.
03
onwardMore search amortized

RL-trained reasoning LLMs

Knowledge and rewarded reasoning patterns are compressed into weights. For familiar problems, the model can produce an effective trajectory without repeating the exploration that shaped its policy.

Across nearly three decades, these systems distribute the work differently while preserving online computation where learned guidance remains incomplete.

Modern reasoning models shift the balance further toward amortization. Reinforcement-learning post-training generates and evaluates trajectories, then incorporates useful regularities into weights rather than storing those trajectories as retrievable episodes. This compression can obscure rare exceptions and leaves weights a weaker guide outside the regions explored during training.

At inference, a model can reopen search for the current problem by extending a reasoning trace, sampling alternatives, invoking tools, or testing actions against the environment. More computation helps only when the system can construct useful candidates, evaluate them reliably, and recognize when a different strategy—or no further search—is warranted.

Part IV

Cognition Coordinates It

Cognition is not a fourth regime beside learning, memory, and explicit search. In the architectural sense used here, it is the organization that makes them operate as one adaptive system.

Each regime offers a different balance of access, cost, and flexibility. Weights provide fast, distributed guidance but do not ordinarily expose the episodes that shaped them; episodic memory can retrieve context-rich precedents but depends on successful indexing and retrieval; explicit search can construct and test novel combinations, but it consumes time and computation.

Greater cognitive capacity is therefore not simply more computation. It is the ability to coordinate these tradeoffs more effectively: to combine prior knowledge with particular experience, recognize what neither resolves, and adapt its deliberation as the situation changes.

Biology offers an analogue, not a blueprint: repeated experience shapes habits, episodic memory preserves particular events, and planning evaluates possible futures. The architectural lesson is that intelligence depends not only on specialized capabilities, but on their coordination.

From philosophy to architecture

Coordination organizes the search process.

That coordination becomes concrete through three operations. Candidate construction determines what possibilities are worth considering. Evaluation supplies the evidence needed to compare them. Control organizes the process—combining learned and retrieved guidance with new search, changing strategy when necessary, and stopping when further work is no longer useful. Learning and episodic memory can improve all three.

A design consequence

Cognition adapts search to the demands of the moment.

One consequence is adaptive computation: novelty, uncertainty, stakes, available resources, and the expected value of further work determine how much explicit search is appropriate. A system may minimize compute while meeting a performance threshold—minimally sufficient search—or maximize performance within a fixed budget, revising its operating point as the situation changes.

Figure 04 · Cognition and the compute–performance frontier

Test-time cognition pushes the compute–performance frontier outward.

Task requirements, current evidence, and resource constraints inform cognition, which coordinates search across learned weights, episodic memory, and explicit search. Selecting mechanisms, allocating effort, and deciding when to redirect or stop are parts of that coordination. Results become new evidence and feed back into cognition, allowing it to readjust. A schematic compute–performance plot shows test-time cognition above and to the left of standard inference scaling. For a required performance level, cognition reaches the target with less computation. For a fixed compute budget, it reaches better performance.

Schematic: the shape and magnitude of the frontier shift depend on the task, metric, and system.
Part V

The Voaige Research Program

The research direction

Building the coordinating layer for more powerful adaptive intelligence.

At Voaige, we treat cognition as a research program rather than a single mechanism. We are studying how learned weights, episodic memory, and explicit search can operate as a coherent system; how the system can adapt as evidence and goals change; and how useful outcomes can become guidance for future decisions.

The program asks:

  • How should the work of search be distributed across learning, memory, and deliberation?
  • What should be generalized into weights, preserved as an episode, or worked out anew?
  • How can these resources compound rather than operate as disconnected modules?
  • How should their coordination adapt with novelty, uncertainty, goals, and experience?

Our first published step is Dynamic Compute Allocation. On Terminal-Bench 2.0 with Mini-SWE-agent, a three-model DCA composite reached 64.8% success at a median cost of $0.26 per task, compared with 57.7% at $0.45 for the GPT-5.2 medium-reasoning baseline. This tests one responsibility of the coordinating layer: adapting models and inference effort as a problem unfolds. The larger aim is to connect that control process with learning and episodic memory so that an agent becomes more efficient now and more capable through experience.

If you are working on adaptive computation, episodic memory, planning, metacognitive control, or inference-time search, we would like to compare notes. Reach us at contact@voaige.com.

Intelligence is search.

Learning carries forward its regularities.

Memory preserves its particulars.

Cognition coordinates it.

That is the research program we are building at Voaige.