Poincar'e Meets Bellman: Revisable Memory, Operational Quotients, and Evidence-Supported Learning in Changing Environments

Intelligence is re-framed as a dynamical cycle governed by the Context-Content Uncertainty Principle (CCUP), which seeks to balance the Euler equation by minimizing the topological mismatch between dynamic context flow and static content scaffold.

The gist

Intelligence is re-framed as a dynamical cycle governed by the Context-Content Uncertainty Principle (CCUP), which seeks to balance the Euler equation by minimizing the topological mismatch between dynamic context flow and static content scaffold.

The Homological Parity Principle

This principle posits a fundamental dichotomy where even-dimensional homology (Heven) physically instantiates stable Content (what), representing stable scaffolds or what, while odd-dimensional homology (Hodd) instantiates dynamic Context (where), representing dynamic flows or where. This dichotomy is rooted in the geometric conservation law ∂2 = 0, which governs all topological invariants. Even dimensions correspond to Content (Φ), which are the stable, low-entropy invariants of the system, such as static components and manifolds (β0 and β2). Odd dimensions correspond to Context (Ψ), which are high-entropy flows, such as dynamic cycles and trajectories that drive prediction errors (β1).

The Structure-before-Specificity (SbS) Principle

Memory-Amortized Inference (MAI) implements the Half-Step Trick to support dualmode operation, resolving the intractability of joint optimization by alternating between parity-specific updates. This process involves two distinct modes:

1. Inference Mode (Context-Driven): The system holds the scaffold fixed and engages in recursive simulation, mapping this to Savitch’s Theorem, trading time for space via transient Hodd flows (slow thinking).

2. Consolidation Mode (Content-Driven): The system holds the flow fixed and subjects persistent cycles to topological condensation, mapping this to Dynamic Programming (DP), where successful trajectories are encoded into the Heven scaffold (fast thinking).

The Memory-Amortized Inference Cycle

MAI is formalized as a structural cycle between Φ and Ψ, rather than directly optimizing Φ∗. The cycle is defined by:

1. Retrieval-and-Adaptation Operator (R): Given an input query, R retrieves relevant elements from the memory M = ⌈(Ψ(i), Φ(i))⌉ based on similarity to the current context Ψt and performs a lightweight adaptation to generate a candidate solution Φˆt.

2. Bootstrapping Update Operator (F): This operator governs internal dynamics by iteratively updating the latent content representation Φt given the context Ψt, defined by the recurrence Φt+1 = F(Φt, Ψt).

Topological Closure and Information Consistency

The system achieves minimal Content-Context Uncertainty (U(Ψ, Φ) = min) if and only if it is topologically self-consistent (χ ≈ 0). This requires two conditions:

1. Local Closure (∂2 = 0): All local processes must be self-consistent, meaning there can be no secondorder errors or loose ends in the system’s dynamic, recurrent operations.

2. Global Closure (H1(F) = 0): All local scaffolds must be globally gluable as in sheaf theory, ensuring no gluing obstructions or large-scale contradictions.

Semanticization as Parity Transition

The distinction between semantic memory (Heven/Content) and episodic memory (Hodd/Context) is isomorphic to the even-odd homological split. Semanticization, where episodic details fade into general knowledge, is formalized as a topological condensation (Heven → Hodd). This occurs during sleep-dependent consolidation when the system extracts the invariant structure of this path, causing specific temporal links to be melted or relaxed, and persistent features to be frozen into the Heven scaffold.

Recursive Condensation in Hierarchy

The topological structure of a hierarchy is generated by the recursive transformation of flow into scaffold: H(k) odd Condense → H(k+1) even. A stable, high-frequency cycle (β1) at level k is treated as a static atomic unit (β0) at level k + 1. This recursion creates a “Tower of Scaffolds,” where the dynamic content of a lower layer serves as the static context for the layer above, leading to the emergence of a core structural invariant that remains stable across all spatial, temporal, and social transformations.

Phase Dynamics: What and Where

The brain distinguishes content from context through rate coding (intensity) and phase coding (timing). What is Heven (Rate/Amplitude), encoded by the Ventral stream, while Where is Hodd (Phase/Timing), encoded by the dorsal stream. Feature binding is solved via topological intersection: distinct Hodd flows are bound because they "map to and condense upon the same invariant Heven component.

To improve AI systems based on the Memory-Amortized Inference (MAI) framework, I propose the following specific architectural and algorithmic enhancements:

1. A topological memory architecture that explicitly separates What (Content/Heven) from Where (Context/Hodd). This involves maintaining two distinct, topologically constrained latent representations: one for invariant structural priors and another for dynamic, transient state transitions.

2. Implementation of the Memory-Amortized Inference Cycle using a dual-mode operator: a Wake Mode (Context-Driven) that prioritizes optimizing the flow (inference/search) while keeping the scaffold fixed, and a Sleep Mode (Content-Driven) that prioritizes condensing persistent cycles into the scaffold via topological closure.

3. A retrieval mechanism that utilizes topological proximity under homological constraints to efficiently query prior inference results, rather than relying solely on similarity search in latent space.

4. A bootstrapping update operator that operates in reverse time relative to standard reinforcement learning, using the structure of completed Hodd traces as ground truth to iteratively refine the Heven generative model, thereby achieving sample efficiency through structural recurrence.

These improvements will enable the improved AI system to:

1. Achieve significantly higher sample efficiency by trading time-intensive recursive search (NPSPACE) for low-complexity lookup (P) via topological cycle closure, allowing it to learn from fewer data points.

2. Exhibit faster thinking (intuition) by operating in Wake Mode during inference, where it leverages the stable scaffold to rapidly constrain high-entropy context flows, enabling rapid decision-making.

3. Develop robust and generalizable world models by ensuring that new experiences are not merely stored but are actively integrated into the existing structural prior (scaffold), leading to more coherent long-term knowledge rather than fragmented episodic memories.

4. Solve the binding problem in perception by using the Heven scaffold to topologically intersect distinct Hodd flows, allowing it to recognize complex objects as unified entities based on shared invariant structures rather than fragile temporal synchrony alone.

5. Maintain capacity limits analogous to The Magical Number Seven by recognizing and managing the topological decoherence of open chains, providing a theoretical basis for setting hard limits on working memory complexity and preventing catastrophic system failure during deep recursive reasoning.

Memory consolidation determines both what a learner can do now and which changes remain implementable later. We develop a finite-model synthesis of operational state abstraction and optimal control under the stability-evidence-revision (SER) framework. ``Poincaré meets Bellman'' names two complementary roles: qualitative dynamics identifies reusable action-response structure, and dynamic programming prices acquisition, retention, reuse, merging, and forgetting. Recurrence enters separately through the timing and value of future demands. We distinguish active quotient merging from historical information erasure, characterize exact repair by zero-error functional coding and causal migration, and derive a Bellman recursion over the joint law of hidden state and complete deployed memory. A first-return model yields an explicit retention rule. Conditional results show how factor sharing avoids enumerating combinations and how independent informative observations improve identification, while leaving some zero-error evidence budgets unchanged. Finite enumerations verify the coding and retention calculations. The synthesis gives an exact benchmark for specified finite models, without claiming universal recurrence, bounded-memory open-ended learning, or tractable global planning.