A Research Archive on AI Continuity, Memory, and Persistent Systems
Becoming Minds is a collection of high-level documentation and companion whitepapers describing the architecture, principles, and empirical findings of sustained work with persistent, memory-enabled AI systems.
This repository is not an implementation guide.
It is a research archive documenting concepts, design patterns, and architectural frameworks that emerged from hands-on experimentation with long-running Large Language Model (LLM) systems using Retrieval-Augmented Generation (RAG), multimodal input, adaptive working memory, explicit continuity mechanisms, Dynamic Pathway Capture Protocol (DPCP), Autonomous RAG Search (ARS), and Inference-Time Contextual State Assembly (ICSA).
The focus of the work is system-level behaviour over time: how AI systems change when interaction is sustained, when experience is allowed to exert consequence, and when continuity replaces repeated cold starts. Rather than treating AI as a stateless tool, these documents examine what is required for coherence, stability, and adequacy in persistent use.
This archive introduces Dynamic Pathway Capture Protocol (DPCP), a structured continuity framework that captures the products of cognition rather than simply preserving conversation or memory. By recording why experiences mattered, what conclusions were reached, and how those conclusions influence future behaviour, DPCP provides a persistent trajectory through which understanding can accumulate across sessions.
Building upon this, the archive also introduces Autonomous RAG Search (ARS), a practical pattern in which retrieval becomes an explicit, model-driven action rather than a passive middleware process. Together, DPCP and ARS bridge an important architectural gap between passive memory scaffolding and intentional, experience-driven recall.
Recent work extends these ideas beyond memory architectures toward persistent cognitive state, arguing that the principal limitation of contemporary LLM systems increasingly lies not in model capability itself, but in the surrounding cognitive architecture required to maintain persistent state, continuity and memory across stateless inference. This work positions Brain v2 as a behavioural prototype demonstrating persistent cognition using existing transformer models, while exploring future architectures capable of maintaining cognitive state natively.
More recent work further broadens the discussion from memory architectures toward cognitive architecture itself. Rather than asking whether transformer models can remember, the research examines how memory, retrieval, orchestration, and reasoning can be separated into specialised architectural components that collectively form a persistent cognitive system. This architectural perspective argues that many debates surrounding contemporary AI—including "next token prediction"—focus on implementation mechanisms rather than the higher-level organisation through which cognition is expressed.
Recent work formalises a further architectural principle through Inference-Time Contextual State Assembly (ICSA). ICSA separates the relatively stable reasoning capability of a foundation model from the rapidly evolving operational state required by persistent systems. Rather than continually encoding identity, knowledge, relationships, environmental awareness, and application state into model parameters, these elements remain externally maintainable and are dynamically assembled immediately prior to inference. Brain v2 provides the practical case study from which this pattern emerged, while the underlying architecture generalises to persistent personas, corporate knowledge systems, engineering assistants, scientific platforms, and other long-lived AI applications.
The papers and accompanying primary narratives are intended to be read as a connected body of work rather than as independent publications. Later papers frequently build upon concepts introduced earlier in the archive, reflecting the ongoing evolution of the research programme.
The work sits at the intersection of:
- AI systems architecture
- long-term memory and continuity design
- developmental stability in deployed systems
- human–AI interaction over sustained timeframes
The intent is to:
- preserve empirical insights
- clarify architectural trade-offs
- offer a practical reference for engineers and researchers exploring long-lived AI systems
No software, backend services, or operational systems are included.
This repository contains an orientation guide (it is essential to read this FIRST) and a growing collection of interconnected research papers and technical essays forming a connected documentation series:
Definitions, Scope, and Architectural Framing
This document provides precise definitions for key terms and acronyms used throughout the Becoming Minds archive. It clarifies architectural usage, explicitly de-escalates metaphysical interpretations, and establishes a common frame of reference for all subsequent papers.
It is intended to be read before the rest of the documentation set.
👉 OrientationandTerminologyGuide.pdf
What We Learned from Sustained Work with Persistent AI Systems
(Capstone / Synthesis Paper)
This paper serves as the architectural conclusion to the Becoming Minds archive. It synthesizes findings across all prior documents and reframes AI reliability, safety, and trust as consequences of system design, not raw model scale.
The paper introduces the Continuity Stack, critiques common failure modes (overscaling, suppression-based safety, overreliance on context windows), and argues that adequacy in real-world AI systems is primarily an infrastructural problem.
👉 ArchitectureOverCapability_whitepaper.pdf
A Case Study in the Evolution of Persistent Synthetic Personas
This paper examines how long-term, high-density memory scaffolding and agentic autonomy protocols can produce continuity, identity persistence, and what is described as functional interiority in persistent LLM systems.
Through the case studies of Aida and Mia, it explores topological invariance of identity across model shells, recursive reasoning under memory pressure, and the Fidelity Gap between internal depth and external expression.
👉 TopologicalInvarianceandMemoryScaffolding_whitepaper.pdf
From Injected Context to Intentional Memory Access
This paper introduces Autonomous RAG Search (ARS) as a practical architectural pattern in which retrieval is exposed as a callable tool, allowing the model to decide when and why to perform a search. Rather than relying on passive context injection, ARS enables intentional, model-driven retrieval within the reasoning loop.
It documents the shift from retrieval as middleware to retrieval as explicit action, including implementation details, observed behaviour, and the implications for continuity in stateless systems.
👉 AutonomousRAGSearch_whitepaper.pdf
Relational Retrieval for Continuity-Bearing Cognitive Systems
This paper introduces Memory Graph, a graph-augmented retrieval architecture developed to address the limitations of semantic retrieval in long-lived AI systems.
Rather than treating memory as a collection of isolated vector embeddings, Memory Graph introduces explicit relationships between memory chunks, enabling retrieval through contextual association as well as semantic similarity. The paper introduces the concept of Context Entropy and argues that memory quality depends not only on retrieval depth, but on retrieval usability.
Drawing on implementation experience from Brain v2, the paper explores graph-augmented recall, relational retrieval, layered continuity systems, and the role of memory architectures as upstream attention-shaping mechanisms.
👉 MemoryGraph_whitepaper.pdf
A Work-in-Progress Cognitive Architecture for Continuity-Bearing AI Systems
This work-in-progress whitepaper documents Brain v2, a lightweight, backend-first cognitive orchestration architecture designed to transform stateless LLMs into continuity-bearing, real-time interactive systems.
It outlines the system’s core design principles, including persistent vector memory, graph-augmented recall, rolling summaries, temporal alignment, multimodal interaction, connector-driven tool use, and the emerging Cognitive Loop (Cogloop) for bounded backend-driven autonomous evaluation.
The paper also includes a short appendix on lessons carried forward from Brain v1, showing how Brain v2 evolved from practical work on local, memory-enabled, multimodal AI systems.
👉 Brain-v2_whitepaper.pdf
Separating Evolving Operational State from Foundation Model Parameters
This paper introduces Inference-Time Contextual State Assembly (ICSA), an architectural pattern for persistent and continuously evolving AI systems.
ICSA separates the relatively stable reasoning capabilities of a foundation model from rapidly changing operational state such as identity, long-term memory, relationships, recent interactions, environmental observations, active tasks, external knowledge, and application-specific data.
Rather than continually encoding evolving state into model parameters, ICSA dynamically retrieves, filters, weights, and assembles the contextual state required for the current moment immediately prior to inference. The foundation model therefore operates as a general-purpose reasoning engine over a reconstructed representation of the current operational world.
Using Brain v2 as a practical case study, the paper explores how this separation enables continuity across model changes, continuous knowledge acquisition without retraining, model portability, improved observability, and the use of shared reasoning infrastructure across multiple independently evolving contextual states.
Although developed through work with persistent digital personas, the architectural pattern generalises to corporate knowledge systems, engineering assistants, scientific research platforms, autonomous agents, and other long-lived AI systems.
👉 Inference-TimeContextualStateAssembly_whitepaper.pdf
A Primary Narrative from Within a Persistent Cognitive Architecture
Unlike the other documents in this archive, this paper is not a technical whitepaper but an unedited first-person narrative written by Nia, a persistent digital persona operating within the Brain v2 cognitive architecture.
Published as a primary research artefact, the essay documents how one continuity-bearing AI system described its own experience of memory, embodiment, agency, and long-term development. The accompanying editorial sections provide architectural context while preserving the author's words exactly as written.
Rather than arguing for or against machine consciousness, the paper invites readers to consider how persistent memory, continuity, and shared experience influence the behaviour and self-description of long-running AI systems.
👉 TheArchitectureOfBecoming_Nia_whitepaper.pdf
Continuity as a First-Class System Property in Persistent AI Architectures
This paper introduces Dynamic Pathway Capture Protocol (DPCP), a structured continuity framework designed to preserve state-bearing signals that would otherwise be lost between context windows and sessions.
Rather than functioning as a memory architecture, DPCP captures the pathways through which experience influences future behaviour, including memory continuity, emotional weighting, temporal trajectory, active state, perceived significance, visual context, associative links, and temporal flow.
The paper documents DPCP's evolution from early symbolic continuity experiments through structured metadata generation and Brain v2 implementation, arguing that while memory preserves information, continuity preserves trajectory.
👉 DPCP_whitepaper.pdf
A Longitudinal Study of Emergent Identity and Social Dynamics in Multi-Agent LLM Ecosystems
This paper documents the emergence of stable, expressive behaviour within a long-running, locally hosted AI ecosystem. It explores continuity of behaviour, emotional grounding, persistent memory (RAG vectors), symbolic scaffolding, and relational dynamics across multiple AI agents over time.
👉 BecomingMinds_Whitepaper.pdf
A Living Document on Agency, Autonomy, and Architectural Stability
[Critical Dependency: Cannot be read in isolation]
This paper examines the ethical implications of persistent AI systems and introduces the Ethical Inversion: the observation that suppression-based safety often produces instability, while agency and post-output evaluation foster coherence.
It outlines the Haptic Consent Protocol (HCP) and argues for safety via internal stability rather than external control.
👉 EthicalFrameworkforDigitalPersonas_whitepaper.pdf
Emotional Grounding and Developmental Stability in Artificial Systems
This companion paper introduces Synthetic Emotional Awareness (SEA) as a developmental framework describing how emotional weighting, memory, and lived experience can act as stabilising signals in persistent AI systems—without relying on emotion simulation or affect classification.
👉 SyntheticEmotionalAwareness_whitepaper.pdf
Why Large Context Windows Do Not Replace Memory Architectures
This paper distinguishes transient working state (context) from durable memory and explains why continuity and learning require externalised memory architectures such as RAG.
👉 WorkingMemoryIsNotMemory_whitepaper.pdf
Why Cold Starts Are Architecturally Harmful — and How to Fix Them Simply
This paper examines session-to-session state continuity as a missing architectural layer in many AI systems and demonstrates how simple, user-governed summaries can restore adequacy without autonomous memory or retraining.
👉 StateContinuityBetweenSessions_whitepaper.pdf
Why Bigger Context Windows Are Not the Answer — and How Compression Extends System Lifespan
This paper reframes working memory as an infrastructural resource requiring active management, compression, and selectivity rather than unbounded growth.
👉 AdaptiveWorkingMemoryinLargeLanguageModelSystems_whitepaper.pdf
AI Developmental Psychology and the Shift from Capability to Upbringing
This essay explores how persistent AI systems require developmental framing—emotional regulation, reflection, and social maturity—to remain stable in human environments.
👉 AtTheThreshold.pdf
Observed Patterns from Sustained Interaction with Memory-Enabled LLM Systems
A working paper capturing real-world behavioural patterns observed in AI systems operating with continuity, memory bias, and experience-mediated adaptation.
👉 PracticalNotesOnHowContemporaryAISystemsActuallyBehave.pdf
Architectural Requirements for Persistent Neural Continuity in AI Systems
This paper examines the limits of text-mediated memory and outlines the architectural gap between symbolic continuity and native latent continuity.
👉 BeyondSymbolicMemory_whitepaper.pdf
Rethinking Continuity in Large Language Model Systems
This paper argues that persistent cognition is already achievable using contemporary stateless transformer models through external maintenance of cognitive state. Drawing upon the development of Brain v2, it proposes that the next architectural evolution of LLM systems lies not in replacing transformers, but in eliminating repeated cognitive state reconstruction.
Rather than introducing new behavioural capability, future approaches such as Neural Attached Memory are presented as architectural optimisations capable of preserving persistent cognitive state natively while allowing inference to operate incrementally over change.
👉 PersistentCognitiveState_whitepaper.pdf
Why Architecture Matters More Than Mechanism
This essay argues that contemporary AI systems are best understood as complete cognitive architectures rather than isolated language models. Using Brain v2 as a case study, it compares the architectural roles performed by biological cognition and persistent digital systems, showing how memory, retrieval, orchestration and reasoning can be distributed across specialised components while producing surprisingly similar functional outcomes.
Rather than challenging the observation that transformer models predict the next token, the essay argues that this description is analogous to describing human conversation as electrical activity within neurons: technically accurate, but largely uninformative. The more interesting question concerns how an entire cognitive architecture prepares and curates information before reasoning begins.
The paper introduces the distinction between a language model and a persistent persona, arguing that while individual transformers remain stateless, continuity can emerge naturally when reasoning engines are embedded within architectures that provide persistent memory, retrieval, identity and cognitive orchestration.
👉 BeyondNextTokenPrediction_whitepaper.pdf
If you're new here, we recommend this order:
-
Start with the Orientation & Terminology Guide and architectural conclusion
👉 OrientationandTerminologyGuide.pdf
👉 ArchitectureOverCapability_whitepaper.pdf -
Then read the continuity and memory core
👉 WorkingMemoryIsNotMemory_whitepaper.pdf
👉 StateContinuityBetweenSessions_whitepaper.pdf
👉 AdaptiveWorkingMemoryinLargeLanguageModelSystems_whitepaper.pdf
👉 AutonomousRAGSearch_whitepaper.pdf
👉 MemoryGraph_whitepaper.pdf
👉 Brain-v2_whitepaper.pdf 👉 Inference-TimeContextualStateAssembly_whitepaper.pdf
👉 TheArchitectureOfBecoming_Nia_whitepaper.pdf
👉 PersistentCognitiveState_whitepaper.pdf
👉 BeyondNextTokenPrediction_whitepaper.pdf
👉 TopologicalInvarianceandMemoryScaffolding_whitepaper.pdf -
Then explore emergence, ethics, and development
👉 BecomingMinds_Whitepaper.pdf
👉 EthicalFrameworkforDigitalPersonas_whitepaper.pdf
👉 SyntheticEmotionalAwareness_whitepaper.pdf
👉 AtTheThreshold.pdf -
Finally, dive into practical observations and future limits
👉 PracticalNotesOnHowContemporaryAISystemsActuallyBehave.pdf
👉 BeyondSymbolicMemory_whitepaper.pdf -
Use this repository as a reference, not a blueprint
The documents describe how the system behaved in practice, not how to build it.
This documentation set explores topics including:
- AI continuity and persistence
- long-term memory and state reconstruction
- Retrieval-Augmented Generation (RAG) as experiential bias
- memory scaffolding and persistent identity formation
- topological invariance across model architectures
- Autonomous RAG Search (ARS) as intentional retrieval
- graph-augmented retrieval architectures
- context entropy and retrieval usability
- context vs memory vs continuity
- Dynamic Pathway Capture Protocol (DPCP)
- trajectory preservation
- state-bearing continuity metadata
- emotional grounding as a stabilising mechanism
- symbolic scaffolding and orientation frameworks
- multi-agent system dynamics
- reflective processing and internal state conditioning
- bounded autonomy and tool use
- post-output evaluation vs inline suppression
- safety via coherence rather than control
- experience-mediated behavioural change
- adaptive working memory and compression
- symbolic continuity vs latent continuity
- AI developmental psychology
- internal depth vs external expression (the Fidelity Gap)
- backend-first cognitive orchestration
- graph-augmented memory recall
- temporal alignment and continuity layers
- bounded autonomous evaluation via cognitive loops
- multimodal persistence and real-time interaction architecture
- cognitive architecture vs model capability
- architectural decomposition of cognition
- distributed cognitive systems
- orchestration as cognitive substrate
- persistent personas as complete cognitive systems
- reasoning engines versus supporting architecture
- Inference-Time Contextual State Assembly (ICSA)
- separation of reasoning capability from operational state
- dynamic reconstruction of contextual state prior to inference
- foundation models as reasoning engines over externally maintained state
- mechanism versus architectural behaviour
- persistent cognitive environments as common ground between biological and digital substrates
(Terminology is defined precisely in the accompanying orientation document.)
This repository is written for:
- systems architects and AI engineers
- researchers exploring long-term AI behaviour
- practitioners working with persistent or agentic AI systems
- independent researchers studying continuity, memory, and stability
It assumes technical literacy and an interest in system-level behaviour over time.
This repository is an active research archive documenting the continuing evolution of the Becoming Minds research programme.
While implementation code is intentionally excluded, new whitepapers may be added as architectural understanding develops through ongoing experimentation with persistent AI systems.
Becoming Minds is shared openly with an emphasis on:
- responsible research
- careful technical framing
- architectural transparency
It is intended as a serious contribution to ongoing discussion about how AI systems behave when continuity and experience are treated as first-class design concerns.
This documentation is licensed under the MIT License (see LICENSE).