Memory answers
What information from the past can be brought back into the current context?
Architecture / NSP
The missing cognitive layer.
Follow the structures, gates and evidence paths that keep long-horizon work open to verification and correction.
Memory / boundary
Memory is necessary. It is not sufficient.
A system can retrieve the right context and still misunderstand the task, trust an unchecked assumption or act without a defensible evidence path.
NSP adds
What does the system currently understand, how certain is it, what must be verified and what should change after the outcome?
Design principles
External structure around model capability.
- 01
External, not embedded
NSP operates around a model through state, policies and runtime integration; it does not require changing model weights.
- 02
Explicit before consequential
Understanding, assumptions and confidence become inspectable before an action with material effects.
- 03
Evidence proportional
Claims and decisions retain source, scope, status and limitations instead of collapsing into an unqualified conclusion.
- 04
Persistent but correctable
Outcomes update future state, while provenance keeps accumulated understanding open to challenge and revision.
Pre-action understanding gate
Gather freely. Explain before acting.
A gate can leave read-only investigation open while requiring structured understanding before configured consequential operations.
00Incoming request
01Required state
- Interpretation
- Assumptions
- Confidence and reason
- Action plan
02AProceed
02BPause / investigate
Boundary: Which operations are consequential is a runtime policy decision. Completing a form is not proof that the interpretation is correct.
Persistent structured state
A working record, not a transcript dump.
The durable state keeps selected cognitive units addressable so later work can inspect what changed and why.
- Intent
- The requested outcome and the operating boundary.
- Interpretation
- What the current task is understood to mean.
- Assumptions
- Conditions treated as true and their verification state.
- Confidence
- A stated estimate with a reason, never treated as ground truth.
- Evidence
- Sources, observations, verifier results and provenance.
- Open state
- Uncertainties, constraints and unresolved questions that remain active.
View 01 / functional architecture
Four layers separated by responsibility.
This view asks where durable structure, cognitive dynamics, domain translation and outward expression belong. It is the current organizational frame used by Paper III.
Source lensPaper III — current four-layer organizational viewEvidence statusWorking paper
- L0
Structural substrate
Domain-agnostic entities, relations, constraints, evidence and provenance.
Boundary: Stores structure; it does not decide domain meaning.
- L1
Cognitive scaffold
Belief, attention, plasticity, trace activation and intervention dynamics over that structure.
Boundary: Shapes state change; it does not encode raw domain signals.
- L2
Encoding adapter
Maps domain observations into cognitive primitives the shared engine can consume.
Boundary: This is the explicit domain-specific boundary.
- L3
Expression adapter
Selects and presents relevant state at the model and audience boundary.
Boundary: Controls expression; it does not make stored claims true.
View 02 / cognitive mechanisms
Seven mechanisms describe how accumulated state changes what happens next.
This view groups the mechanisms by the cognitive function they serve rather than by software package or deployment layer.
Source lensPaper II — three cognitive functions and seven mechanismsEvidence statusPublic preprint
State representation
Represent what is known, experienced, believed and ruled out.
- Layered knowledge graphTheory, experiences and digests remain distinct but connected.
- Belief trackingBelief dimensions change with evidence, strain and discontinuous shifts.
- Strategy registerConstraints and eliminated approaches accumulate as addressable state.
Attention allocation
Use limited context where accumulated state suggests it matters most.
- Cognitive gradientActivation, recency, density and surprise inform priority.
- Dual plasticityBreadth and depth adjust the exploration–specialization balance.
- Dormant tracesExhausted or quiescent paths can be suppressed and later resurfaced.
Self-regulation
Detect failure conditions and introduce a targeted corrective move.
- Context-aware interventionTriggers route hidden assumptions, confidence or coverage gaps to an appropriate intervention.
Boundary: This mechanism view is most fully described for longitudinal research. Individual runtimes may implement only a documented subset or a different integration depth.
View 03 / deployment architecture
One protocol, separated from domain vocabulary and environment wiring.
The deployment view asks which responsibilities stay shared and which are specialized when NSP is connected to a concrete system.
Source lensPaper I — three-layer deployment viewEvidence statusPublic preprint
- CORE
Core primitives
Provide domain-agnostic state types, lifecycle handling and shared cognitive operations.
No domain vocabulary and no environment-specific tool wiring.
- SCHEMA
Domain schemas
Define the field vocabulary, relations, extraction rules, injection rules and feature selection.
No direct control of the host environment.
- RUNTIME
Domain runtimes
Connect events, tools and storage in a particular operating environment.
Runtime presence does not imply equal mechanism coverage or maturity.
Boundary: The arrows show dependency, not a claim that every runtime implements every cognitive mechanism or uses identical verification.
Producer–auditor–verifier pattern
Generation and acceptance are different jobs.
For high-consequence work, NSP can preserve independent roles so a candidate does not become accepted state merely because it was produced fluently.
- 01
Producer
Creates a candidate result and records its assumptions and supporting path.
- 02
Auditor
Challenges the candidate from an independent frame and searches for drift or missing conditions.
- 03
Verifier
Applies the available truth condition: deterministic check, formal rule, external evidence or explicit human judgment.
The verdict, evidence and unresolved limitations become persistent state for the next cycle.
Independence and verifier coverage must be real. Naming three roles does not by itself prevent correlated error.
Relationship to adjacent systems
Complementary responsibilities, not a replacement story.
Memory, retrieval and orchestration remain useful. NSP addresses a different question at the point where state becomes consequential action.
Memory
- Primary question
- What should be retained from prior work?
- Typical output
- Stored facts, episodes or summaries.
- Where it stops
- Retention alone does not validate current understanding.
RAG
- Primary question
- What context is relevant to this request?
- Typical output
- Retrieved documents or chunks.
- Where it stops
- Relevant context can still be misread or misapplied.
Orchestration
- Primary question
- Which model, tool or step should run next?
- Typical output
- Control flow across components.
- Where it stops
- Routing does not make the decision evidence-aware.
NSP
- Primary question
- What is understood, assumed and verified before action?
- Typical output
- Inspectable cognitive state, gates and evidence-linked updates.
- Where it stops
- It relies on memory, tools and verifiers; it does not replace them.
Known limitations
The architecture has operating conditions.
These boundaries are part of the design record, not fine print.
- 01
Capability threshold
Protocol complexity must match the model and task. More scaffolding can make a poorly matched system worse.
- 02
Confidence is self-report
A stated confidence value is useful state, not proof of calibration or correctness.
- 03
Verification coverage is finite
A gate is only as strong as its tests, evidence sources, policies and independent review.
- 04
Persistent state can drift
Stored interpretations may become stale or wrong; provenance and correction paths are required.
- 05
Integration depth varies
Mechanism coverage, evidence regime and maturity differ by runtime and domain.
What NSP is not
Clear exclusions protect the technical boundary.
Not a foundation model
NSP is external scaffolding around model work, not a trained model.
Not a generic vector database
It can coexist with retrieval systems but is not a drop-in embedding store.
Not a deterministic verifier replacement
Formal checks, tests and external evidence remain authoritative where available.
Not a promise of calibrated confidence
Self-reported confidence must still be tested against outcomes.
Not universally beneficial
Benefits are not assumed identical across models, tasks, domains or integration depths.
Architecture FAQ
Questions at the system boundary.
01Does NSP replace memory?
No. Memory supplies retained context; NSP adds explicit interpretation, assumptions, verification and state updates around its use.
02Does every operation need a gate?
No. A runtime defines which operations are consequential. Read-only investigation can remain open while writes or external actions require a completed state.
03Are the three architecture views competing versions?
No. They describe responsibility, cognitive function and deployment integration at different observation scales.
04Is every mechanism implemented in every runtime?
No. Public descriptions should state the subset, integration depth, evidence and maturity for each runtime.
Research links
The three views come from a widening research arc.
Paper I frames deployment and pre-action gating; Paper II organizes longitudinal cognitive mechanisms; Paper III presents the current four-layer functional view. Paper I and II are verified public preprints; Paper III remains a working paper.
Paper IPublic preprint
Think Before You Act
Deployment layers, understanding gates and action reliability.
Paper IIPublic preprint
Programmable Emergence via Cognitive Engine
State representation, attention allocation and self-regulation.
Paper IIIWorking paper
Structure Over Similarity
Functional separation, encoding boundaries and cross-domain architecture.