xLMP:
The AI Memory
Control Plane
The forthcoming paper explains why persistent AI memory creates a data-movement problem, why full-context replay is not enough, and how memory, reasoning, and authority must remain separate.
Pre-Release State
The new white paper is not public yet. This page previews the questions it will answer and points visitors to the benchmark evidence already available on the site.
Architecture follows the measurements.
- Why does persistent AI memory become a data-movement problem?
- Why is full-context replay expensive even before physical memory is exhausted?
- Why can indexed retrieval win dramatically in one workload and lose in another?
- How can a system select the lower-work path without exposing the private routing recipe?
- How should persistent evidence remain separate from execution authority?
- How do these principles behave across accelerator architectures?
The Mandate
Autonomous AI systems have execution capability but lack the infrastructure layer beneath it: persistent, portable, verifiable memory that survives models, sessions, devices, and hardware generations. The model context window is working memory. Applications that require durable state must reconstruct it at every invocation, paying the full prefill cost each time, approaching context limits as state accumulates, and losing continuity when models change.
xLMP, the ExergyNet Ledger Memory Protocol, defines the missing layer — the AI Memory Control Plane: the persistent infrastructure that governs which identified, bounded, provenance-bearing evidence enters an agent's computation; where that state survives across models, devices, and accelerators; and how evidence integrity, provenance, and authority are maintained as independent, separately verifiable properties.
Problem
Stateless Intelligence
Most AI systems repeatedly replay history, lose state across model boundaries, and rely on retrieval systems that may omit the exact fact required for high-stakes reasoning.
Solution
Persistent Infrastructure Layer
Preserve state outside the model as integrity-rooted objects, address evidence deterministically with provenance, verify the execution path, and stage only the bounded context required by the current operation.
Four System Invariants
ExergyNet separates persistent memory from temporary context and separates the guarantees of computation, hardware, physical observation, and ledger settlement.
Invariant 01
Memory Is Persistent. Context Is Temporary.
A model should not reread the complete history of a system for every operation. Persistent state remains in the Vault; only the required evidence enters active compute.
Invariant 02
Evidence Must Be Addressable and Complete.
Approximate retrieval may discover candidates. The authoritative execution path must receive the complete object selected for the operation, with provenance and integrity preserved.
Invariant 03
Different Trust Domains Require Different Proofs.
Digital execution, device identity, physical observation, source authenticity, and settlement finality are distinct claims. ExergyNet binds them without treating them as interchangeable.
Invariant 04
State Must Survive Model and Hardware Change.
Persistent memory, evidence lineage, identity, and operational state should remain intact when the model, accelerator, cloud, chain, or network changes.
The ExergyNet Architecture
ExergyNet is not a single chain, model, proving system, or application. It is a layered machine infrastructure in which memory, intelligence, verification, transport, and settlement remain independently addressable but operationally coordinated.
Layer 01
Persistent Memory
Exergy Vault and xLMP preserve long-lived machine state outside active model context.
- Integrity-rooted objects and source provenance
- Bounded evidence retrieval
- Model-independent state continuity
Layer 02
Intelligence & Orchestration
SEI-Vanguard resolves memory, selects evidence, dispatches models and tools, and preserves resulting state.
- Policy-aware execution
- Model and accelerator routing
- Operational receipts and state transitions
Layer 03
Verification & Attestation
Proof receipts, hardware signatures, source attestations, and physical-world evidence form a traceable verification chain.
- Zero-knowledge execution receipts
- Hardware-backed signatures
- Authenticated external data
Layer 04
Transport & Presence
Omega Carrier and the broader communications layer preserve identity, signaling, and machine presence across constrained networks.
- Infrastructure-agnostic signaling
- Intermittent-connectivity operation
- Identity-preserving machine transport
Layer 05
Settlement
Solana and Base provide complementary settlement rails for proof anchoring, fees, governance, and economic finality.
- Multi-rail execution
- Protocol metering
- Operator and treasury settlement
Layer 06
Applications
MyMonitor and other applications consume the underlying memory, verification, transport, and settlement services without redefining them.
- Clinical and scientific systems
- Regulated enterprise agents
- Robotics and sovereign infrastructure
Exergy Vault & xLMP
Persistent Machine Memory Without Proportional Context Growth
Exergy Vault is the persistent memory substrate of ExergyNet. It stores source data, machine state, evidence lineage, and operational outputs as integrity-rooted objects that remain independent of any single model, accelerator, cloud provider, or application.
xLMP is the memory-delivery protocol. It resolves the relevant memory root, selects the smallest complete set of authoritative objects required by the operation, validates their integrity, and stages bounded evidence into active compute.
Context Completeness
Complete Objects, Not Arbitrary Fragments
xLMP preserves the identity and boundaries of the selected object so the model receives the complete authoritative evidence chosen for the operation.
Bounded Compute
Stored Memory Can Grow Without Prompt Replay
Persistent knowledge remains outside active context. Only the evidence required for the current task enters the accelerator.
SEI-Vanguard
The Native Memory and Execution Control Plane
SEI-Vanguard is the orchestration layer that coordinates memory, models, tools, policies, and execution environments. It is not limited to one model family or accelerator class.
For each operation, Vanguard resolves the relevant memory root, identifies the required complete objects, excludes unrelated evidence, validates integrity, dispatches the appropriate model or tool, and preserves the resulting state back into persistent memory.
Vanguard Operational Cycle
Verifiable Execution
Cryptographic Receipts for Defined Computation
ExergyNet can bind a defined computation to committed inputs and outputs through zero-knowledge execution receipts. A valid proof establishes that the specified program executed according to the inputs committed by the proof. It does not, by itself, establish the semantic truth of external data or guarantee correct model reasoning.
Recursive aggregation can compress many independently verified jobs into a bounded settlement artifact, lowering the amortized verification and settlement overhead per operation as aggregation density increases.
Recursive Proof Aggregation
Each job carries its own settlement overhead.
Many verified jobs are represented by a compact recursive artifact.
Aggregation reduces amortized on-chain settlement overhead toward a protocol-defined floor. It does not eliminate proving work, orchestration, or data-availability requirements.
Hardware & Physical Attestation
Proof-of-Exergy, Edge Witness, and Device-Authenticated Evidence
Proof-of-Exergy is one verification primitive inside ExergyNet. It binds device identity, signed execution telemetry, and committed computational work into a traceable receipt. Physical measurements may be attached to that receipt as evidence of environmental or operational state.
Physical telemetry does not replace a computation proof. A secure device signature can establish which authorized hardware produced an observation. A zero-knowledge receipt can establish that a defined computation executed correctly. A ledger anchor can establish ordering and finality. ExergyNet combines these claims while preserving their separate assumptions.
Edge Witness
Hardware-Backed Observation
Secure-device signatures bind sensor or environmental observations to an authenticated hardware identity.
Proof-of-Exergy
Committed Useful Work
Defined computational work, execution telemetry, and resulting commitments are combined into an auditable proof receipt.
AERIS & Authenticated External Data
Bridging Legacy Web Systems Into Verifiable Agent Workflows
Autonomous systems must frequently consume data from web services, enterprise systems, and legacy APIs that were never designed for cryptographic verification. AERIS provides a bridge between those sources and the ExergyNet evidence chain.
The public architecture exposes authenticated source identity, committed response content, retrieval time, and resulting evidence receipts without disclosing private proving mechanics or application credentials.
Source Integrity Flow
Omega Carrier
Sovereign Transport, Presence, and Continuity
Persistent memory and verifiable execution are incomplete without resilient communication. Omega Carrier provides the transport and presence layer through which autonomous systems signal, exchange evidence, and preserve operational identity across changing networks.
The architecture is designed for infrastructure-agnostic signaling and can operate across standard internet transport, direct local links, and constrained or intermittent connectivity environments.
Presence
Persistent Machine Identity
Nodes and agents maintain discoverable operational presence independent of any single messaging provider.
Transport
Multiple Connectivity Paths
Evidence and state transitions can move across internet, local, and edge transports according to availability and policy.
Multi-Chain Settlement
Solana and Base as Complementary Protocol Rails
ExergyNet is designed to operate across multiple settlement environments, including Solana and Base. Protocol functions are assigned according to latency, cost, compatibility, and finality requirements rather than forcing every operation through one chain.
Settlement records may anchor proof receipts, usage metering, node compensation, governance actions, and economic finality. The user-facing service layer remains separate from internal protocol accounting and operator settlement.
Solana Rail
High-Throughput Settlement
Optimized for rapid program interaction, metering, and low-latency protocol coordination.
Base Rail
EVM Compatibility
Supports EVM-aligned integration, application interoperability, and complementary settlement requirements.
Economic Design
ExergyNet separates user-facing service payment from internal protocol accounting. Fees may be settled through supported native assets, while operator compensation and treasury allocation follow protocol-defined rules that may vary by resource class and service level.
Trust Model
The architecture is explicit about what each verification layer proves and what remains outside its guarantee.
| Layer | Primary Guarantee | Does Not Guarantee |
|---|---|---|
| xLMP memory root | Evidence identity, integrity, provenance, and object completeness for the addressed memory. | Correct model reasoning or correct source semantics. |
| Zero-knowledge receipt | The defined program executed according to the committed inputs. | Truthfulness of external data that was committed as input. |
| Hardware signature | An authorized device produced the signed observation or receipt. | That every attached sensor observation was physically accurate. |
| AERIS source receipt | The retrieved response is bound to an authenticated source interaction. | That the external source itself is free from error or manipulation. |
| Ledger anchor | Ordering, timestamping, and tamper-evident finality of the committed record. | Semantic correctness of off-chain computation or source data. |
| Vanguard policy receipt | Evidence selection, tool use, and state transitions followed the declared operational policy. | Absolute model infallibility. |
Measured Validation
Selected benchmark findings demonstrate bounded active context, cross-accelerator portability, compact deterministic retrieval, and the difference between complete-object delivery and top-k fragment retrieval.
Correct-task throughput over full-context replay for the tested workload on the same H200 and model.
Approximate xLMP prompt-token range as stored memory grew 8K→285K tokens on the H200 benchmark workload.
Entity and NOT_FOUND retrieval accuracy at the validated 10 MB compact-index scale.
xLMP accuracy versus 82.6% for the tested sparse top-k RAG path on cross-document synthesis.
positive scaling
Across the tested 32K–10M nominal corpus envelope, mean staged context (K) showed no material positive growth. 10-point fit: b = 2.79×10⁻⁶ (95% CI includes zero; development data). Genuine 10M holdout: 5/5 runs deterministic, 0/190 per-query mismatches, mean K = 895.96.
What the Benchmarks Establish
- Persistent memory can grow while active prompt context remains bounded — observed on H200 (8K–285K corpus) and on the LNES-82C campaign (32K–10M corpus).
- Across the tested 32K–10M nominal envelope, mean staged context showed no material positive scaling with corpus size. Memory Growth ≠ Inference Growth, within the validated envelope.
- The same memory plane can operate across T4, A10, and H200 environments.
- Record-aware deterministic retrieval can remain compact, fast, and integrity-preserving.
- Top-k fragment retrieval can omit independently listed evidence before a model receives the question.
- Complete object delivery reduces that specific pre-model failure mode, while model reasoning remains a separate responsibility.
- All N→K scaling claims are bounded to their tested corpus size, model, and hardware topology and do not extend to arbitrary scales, models, or architectures.
Application Domains
Clinical AI
Persistent Patient and Protocol State
Bounded evidence, source provenance, escalation rules, and verifiable operational history across models and care environments.
Scientific Systems
Traceable Research Memory
Long-lived experimental state, source lineage, reproducible computation, and evidence-preserving agent workflows.
Regulated Enterprise
Auditable Autonomous Operations
Policy-bound agents with complete evidence delivery, tool receipts, and persistent state across vendor boundaries.
Robotics & Sovereign Infrastructure
Memory Beyond the Cloud Boundary
Hardware-attested observation, resilient transport, and persistent machine identity across edge and disconnected environments.
Deployment Status & Roadmap
Architecture descriptions are separated from deployment status so validated capabilities are not confused with planned research.
| Capability | Status | Current Scope |
|---|---|---|
| Exergy Vault and bounded xLMP retrieval | Validated | Persistent storage, bounded evidence delivery, compact deterministic indexing, integrity verification. |
| SEI-Vanguard orchestration | Operational | Memory resolution, model/tool dispatch, state preservation, evidence-aware execution. |
| Solana and Base settlement rails | Operational | Complementary protocol coordination and settlement functions. |
| Hardware and physical attestation | Validated | Device-authenticated evidence and signed execution telemetry. |
| Recursive proving at broad production scale | Research | Proof aggregation, production hardening, cost optimization, and external engineering review. |
| Large-scale compact-index validation | Validated | 1 GB / 100-query narrow-population compact-index validation complete (52,753.8× median paired speedup; 0.72× aggregate in realistic mixed vault — both numbers reported). Results are workload-dependent. |
| N→K corpus scaling (32K–10M) | Validated | Nine-point A100 TP=4 ladder (32K–4M): no material positive K scaling detected; 9-point fit b = 7.83×10⁻⁷ (CI includes zero). Holdouts at 1M/2M/4M all 5/5 deterministic runs, 0/190 per-query mismatches. Gates A and B failed at 500K corpus peak — preserved without relabeling; Gate D PASS. Genuine 10M holdout (5/5 deterministic runs, 0/190 mismatches, mean K = 895.96, WSL2/Windows): K_FLAT_OR_STABLE within established 764–908 envelope; Q_BEND FORMALLY TRIGGERED at 39.5% (0.5pp below 40% floor, Q3/Q4 adversarial density — retrieval-quality finding, not system error); C_R(10M) hardware-confounded. All results bounded to tested model, hardware, and corpus construction. |
Machine Interface
ExergyNet exposes canonical machine-readable manifests, capability descriptors, and integration surfaces for autonomous agent discovery.
llms.txt
LLM Discovery Manifest
Plain-text protocol description for language models and autonomous development agents.
View llms.txt →.well-known/exergynet.json
Protocol Coordinates
Canonical network identifiers, supported capabilities, settlement rails, and gateway metadata.
View JSON →MCP / OpenAPI
Agent Integration Surface
Machine-readable tools, schemas, and SDK pathways for ExergyNet applications and enterprise systems.
View Integration Docs →Public Architecture, Protected Mechanism
ExergyNet publishes system interfaces, architectural invariants, trust assumptions, public receipts, integration schemas, and measured outcomes. The mechanisms that create the operational advantage remain protected as trade secrets or are disclosed only through controlled technical review.
Persistent Memory. Verifiable Execution. Sovereign Operation.
Integrate ExergyNet as the memory, orchestration, verification, and settlement fabric beneath autonomous systems.