Memory as Geometry: Foundations of Coherent, Verifiable, and Private Persistent State

Liquid Memory Series — Expanded Synthesis

July 2, 2026

Memory has traditionally been treated as a secondary concern in the design of intelligent systems — something to be managed through larger context windows, more parameters, or external retrieval mechanisms. This series takes a different starting point: that memory itself can be understood as a structured geometric object, subject to the same principles of curvature, constraint, and verifiability that have proven powerful in mathematics, physics, and information theory.

The ideas here explore what becomes possible when memory representations are required to live on manifolds — particularly hyperbolic ones — when updates respect geometric structure, when memory can be decomposed into modular and hot-swappable verifiable units, and when privacy is treated as a first-class architectural property. The approach draws on information theory, hyperbolic geometry from general relativity, the manifold hypothesis in machine learning, Riemannian optimization, topological data analysis, and modern structured state space models, while remaining grounded in practical concerns of efficiency, coherence, controllability, and agency.

The central claim is not that scale is unimportant, but that structure matters — and that certain kinds of long-term coherence, branching reasoning, verifiable persistence, and agentic behavior become significantly more tractable when memory is given geometric form. These are not offered as a finished architecture, but as principles and patterns that can be combined, extended, or selectively adopted.

The Deep History: Information Theory Meets Geometry

To understand why geometric memory is not just an engineering trick but a deep alignment with how information and reality appear to work, we must look at the historical threads that converge here.

Information Theory: The Birth of a Quantitative View of Knowledge

In 1948, Claude Shannon published "A Mathematical Theory of Communication." At its core was a radical idea: information is not about meaning or semantics but about uncertainty and surprise. Entropy — the expected information content — quantifies the average surprise in a source. Mutual information measures how much knowing one variable reduces uncertainty about another.

Shannon's framework treated communication as a problem of reliable transmission over noisy channels. The key insight was that structure reduces uncertainty. Redundancy, error-correcting codes, and compression all exploit statistical structure in the source.

This was the first rigorous mathematical language for talking about knowledge as something quantifiable and manipulable. It set the stage for thinking about memory not as a passive store but as an active structure that shapes what can be known, retrieved, and acted upon.

General Relativity and Hyperbolic Geometry: Curvature as Structure

Parallel to this, Einstein's general relativity (1915) showed that gravity is not a force but the curvature of spacetime. The geometry is described by the Einstein field equations, where matter and energy tell spacetime how to curve, and curved spacetime tells matter how to move.

Crucially, the mathematics of GR relies on Riemannian geometry — the generalization of Euclidean geometry to curved manifolds. For certain solutions (like anti-de Sitter space in AdS/CFT correspondence), the geometry is hyperbolic.

In hyperbolic space, as we saw in the neuroscience and ML literature, volume grows exponentially. This is not a coincidence for physics either.

The Holographic Principle and Information on the Boundary

In the 1990s, Gerard 't Hooft and Leonard Susskind proposed the holographic principle: the information content of a volume of space can be described by a theory on its boundary, with one less dimension. The Bekenstein-Hawking entropy of a black hole — proportional to the area of the event horizon, not the volume — is the most famous example.

Juan Maldacena's AdS/CFT correspondence (1997) made this concrete: a theory of gravity in anti-de Sitter (hyperbolic) space is equivalent to a conformal field theory on its boundary. Information and geometry are dual.

This is profound for memory: information has a geometric home. The "bulk" (volume) is encoded on the "boundary" (surface). Hyperbolic geometry naturally supports this kind of encoding because of its exponential volume growth — you can pack exponentially more information on the boundary while the bulk remains coherent.

The same mathematics that describes black hole information and the emergence of spacetime from quantum entanglement also appears in modern efforts to embed hierarchical knowledge in AI.

The Manifold Hypothesis in Machine Learning

In the 2000s and 2010s, the "manifold hypothesis" became central to deep learning theory. The idea, articulated clearly by Yoshua Bengio and others, is that real-world high-dimensional data (images, text, sensor streams) actually lies on or near much lower-dimensional manifolds embedded in the ambient space.

If data lives on a manifold, then:

Local operations (like gradients) should respect the manifold's geometry.
Distances and similarities should be measured along the manifold (geodesics), not straight lines through the ambient space.
Representations that unfold or respect the manifold will be more efficient and generalizable.

This is why hyperbolic embeddings took off for hierarchical data: the data's intrinsic geometry is hyperbolic, not Euclidean. Forcing it into flat space requires distortion and waste.

The manifold hypothesis is not just about efficiency. It is a statement about the structure of the world that intelligent systems must discover and exploit. When we give our memory systems the right geometry, we stop fighting the data and start working with its natural shape.

From Physics and Information Theory to Agentic Memory

These threads — Shannon's quantification of information, Einstein's curved spacetime, the holographic encoding of bulk information on boundaries, and the manifold hypothesis — converge on a powerful idea for artificial agents:

Memory should be a first-class geometric citizen.

When memory lives on a manifold (especially hyperbolic), several agent-enabling properties emerge naturally:

Persistent identity across time: The geometry provides a stable "shape" that survives updates.
Branching and planning: Hyperbolic space has natural room for speculative futures and pruning.
Verifiability: Topological signatures (persistent homology) can certify that a memory pack has not drifted.
Hot-swappability: Modular packs can be loaded, merged, or rolled back while respecting geometric constraints.
Efficiency at the edge: Exponential capacity means rich memory in low active dimensions — perfect for local, private agents.

This is what "Liquid Memory" and hot-swappable geometric packs make concrete. It is not an accident that the mathematics of black hole information, hierarchical embeddings, and manifold learning all point in the same direction. They are describing the same underlying phenomenon from different angles: structured information wants to live in curved space.

Part 1: The Geometric Substrate — Why Hyperbolic Space Is Built for Memory

(Expanded from original with history integrated above; now focused on practical geometry for agents.)

Most modern AI systems still represent knowledge as vectors in Euclidean space — flat, uniform, and isotropic. This works for many tasks but fails for hierarchical, branching, long-term memory.

In Euclidean space, volume grows polynomially. Embeddings of deep hierarchies crowd and distort. This is why we need very high dimensions or accept loss of structure.

Hyperbolic geometry solves this because volume grows exponentially with radius. Tree-like structures embed with low distortion. Geodesic distance becomes a natural, meaningful measure of relevance.

The Hippocampus and Hyperbolic Maps (Biological Precedent)

Neuroscience shows the hippocampus organizes spatial and episodic memory using hyperbolic geometry. This is not metaphor — the representational geometry of memory in brains is curved in the same way that supports efficient hierarchy.

Artificial systems can do the same.

Practical Properties for Agentic Memory

Property	Euclidean	Hyperbolic	Benefit for Agents
Volume growth	Polynomial	Exponential	High capacity in low dimensions
Hierarchical/branching data	Crowding	Natural embedding	Coherent long-term + speculative plans
Similarity	Often distorted	Geodesically meaningful	Reliable retrieval and relevance
Compression	High dims needed	Structure-aware	Efficient local deployment

This geometry, combined with selective state-space models (Mamba lineage), gives us the substrate for Liquid Memory — memory that flows and adapts while preserving deep structure.

Basic hyperbolic distance (Poincaré ball) remains useful as a retrieval key.

Part 2: Manifold-Aware State — Making Memory Live on the Geometry

To keep running state on the manifold:

Log map (unwrap): Project to tangent space for linear operations.
Exp map (fold): Retract back to manifold.

For hyperboloid model:

\exp_0(v) = \left( \cosh(\|v\|), \frac{\sinh(\|v\|)}{\|v\|} v \right)

\log_0(x) = \frac{\text{acosh}(x_0)}{\sqrt{x_0^2 - 1}} \cdot x_{1:d}

These operations let recurrent or selective updates happen while the state remains geometrically valid.

Retractions and vector transport (parallel transport approximations) keep updates and gradients consistent after memory injection.

Part 3: Controlled Memory Injection — Gates, Parallel Transport, and Stable Integration

To inject new information without destroying existing structure:

Gated Injection

A learned gate controls movement along the manifold in tangent space, followed by retraction. This is the geometric version of a forget/update gate.

Parallel Transport

For fidelity when memory comes from distant regions of the manifold, transport along the geodesic before integration. This preserves lengths and angles.

Tying to Liquid Mamba

Selective SSM + manifold state + gated (optionally transported) injection = Liquid Mamba.

The selective mechanism chooses what, the manifold protects shape, the gate controls how much, transport ensures respect for structure.

This directly enables agentic memory: persistent, coherent state that supports long-horizon planning, branching, and self-correction without catastrophic forgetting.

Part 4: Modular Knowledge Packs — Hot-Swappable, Compressible, and Verifiable Memory Units (Expanded)

Once memory is geometrically structured, it becomes possible to treat coherent subsets as modular, hot-swappable units — knowledge packs.

What Is a Knowledge Pack?

A self-contained geometric unit containing:

Manifold-resident embeddings (hyperbolic vectors)
Topological signatures (persistent homology, persistence diagrams)
Optional provenance, versioning, and access metadata

Packs can be created, stored, retrieved, versioned, merged, and transferred independently.

Hot-Swappability

Only relevant packs are unfolded into active state. Others remain compressed on the manifold. Geometric and topological constraints act as guardrails during swap-in/swap-out. An agent can "load" a specialized pack for a domain, work with it, then swap it out without corrupting the core state.

Compression and Efficiency

Hyperbolic embeddings are naturally compressive for hierarchical data. The same information occupies lower active dimensionality. Packs can be stored highly compressed and decompressed (unfolded via exp map) only when needed. The active model stays small while the effective memory is large — ideal for edge and long-running agents.

Versioning, Merging, and Rollback

Topological signatures provide verifiable fingerprints:

Detect drift or corruption on load.
Roll back merges that violate expected topology.
Merge along geodesics or controlled interpolations that stay on the manifold.

This turns memory into something closer to version-controlled, auditable modules rather than an opaque vector soup.

Distillation and Training on Packs

Large models can distill experience into compact geometric packs. Smaller agents can then train or adapt directly on these packs. Packs themselves can undergo additional consolidation before re-injection. This creates a clean separation between heavy distillation and lightweight, ongoing agent operation.

Part 5: Liquid Systems in Practice — Putting It All Together

The Core Architecture

Active Engine — Small, efficient selective SSM (Mamba-style) with manifold-resident hidden state. Runs on local/edge hardware.
Memory Store — Library of geometric knowledge packs. Compressed by default, unfolded on demand.
Integration Layer — Gated (and optionally parallel-transported) injection, geometric + topological retrieval, retraction after updates.

Typical Long-Running Agent Flow

Load active manifold state + currently relevant packs.
Evolve state selectively on the manifold.
Process new experience → candidate pack → gated injection.
Branch speculative reasoning on the manifold; score and prune.
Consolidate updated traces back into packs via distillation.
Persist: fold and store. Memory remains as a growing library of verifiable packs.

Manifold constraints and topological invariants provide continuous guardrails.

Why This Enables Agentic Memory

Persistence without bloat: Rich history in compressed packs.
Coherence: Geometry prevents destructive interference between old and new.
Agency: Natural support for branching, planning, self-correction, and long-horizon goals.
Modularity: Hot-swappable expertise without full retraining.
Verifiability: Topological and geometric checks make tampering or drift detectable.
Privacy and locality: Everything can run on-device with selective decryption.

Part 6: Encryption and Verifiable Memory — Keeping the Stack Private and Auditable (Expanded)

Why Encryption Fits the Geometry

Structured representations (manifold embeddings + topological signatures) are modular by nature. This maps beautifully onto encryption boundaries.

Homomorphic Encryption on Geometric Representations

Using schemes like CKKS for approximate arithmetic on real vectors:

Store embeddings encrypted.
Perform similarity search, low-depth additions, and gated scoring homomorphically.
Return results for local decryption and geometric verification.

This enables private nearest-neighbor retrieval, encrypted memory injection scoring, etc.

Practical patterns (2026 state of the art):

Long-term storage: CKKS-encrypted embeddings + separately signed topological metadata.
Fast path: Hash-based retrieval + selective decrypt + geometric re-ranking.
Injection: Homomorphic scoring + trusted execution environment (TEE) for actual gated update + retraction.
Transfer between agents: Fully encrypted packs with signatures.

Verifiable Hot-Swappable Encrypted Packs

A pack can be:

Folded and encrypted at the end of a session.
Stored or transmitted encrypted.
On reload: signature verification + lightweight homomorphic consistency check + geometric/topological validation before any injection.

This makes memory both private by default and auditable.

Synthesis

Treating memory as a first-class geometric and topological object — living on hyperbolic manifolds, injected in controlled ways, decomposed into hot-swappable verifiable packs, and protected by encryption and topology — delivers coherence, efficiency, modularity, and agency that flat vector memory cannot match.

The history from Shannon's information theory through Einstein's curved spacetime, the holographic principle, the manifold hypothesis, and modern geometric deep learning all point to the same conclusion: information has a preferred geometry. When we give memory systems that geometry, agentic behavior stops being an add-on and becomes a natural consequence.

The tools exist. The mathematics is sound. The work now is careful implementation, rigorous verification, and real-world experimentation with long-running, local, private agents.

This post expands the Liquid Memory series, integrating history from information theory and general relativity, the manifold hypothesis, and the enabling of agentic memory through hot-swappable geometric packs.