<img src="https://r2cdn.perplexity.ai/pplx-full-logo-primary-dark%402x.png" style="height:64px;margin-right:32px"/>

Research Framework: From Functional Persona to Narrative Character in Multi-Agent AI Systems

Software Development Pathway

Architecture Extension Map

Your existing Dual-Session Unifier model reveals an architecture explicitly prepared for narrative character expansion. The system separates functional execution (primary Gemini agent) from reflective interpretation (Claude-powered Unifier), creating a natural site for character emergence.¹

Implementation Sequence:

1. Persona Prompt Library Structure (/src/mia-code/src/personas/)
   • Create modular system prompt files as TypeScript constants
   • Each character exports: CHARACTER_SYSTEM_PROMPT, CHARACTER_BACKSTORY, CHARACTER_VOICE_PATTERNS
   • Example structure: skeptical-engineer.persona.ts, whimsical-storyteller.persona.ts
2. Configuration Schema Evolution (/src/mia-code/src/config.ts)

interface MiaCodeConfig {
  activePersona: string;
  personaMemoryDepth: 'session' | 'persistent' | 'ephemeral';
  narrativeCoherence: 'strict' | 'flexible';
}

3. Dynamic Prompt Selection Engine (/src/mia-code/src/unifier.ts)
   • Implement PersonaRegistry class with character lookup
   • Add loadPersonaPrompt(personaId: string) function
   • Integrate session-aware context injection before Claude invocation
4. CLI Surface Expansion (/src/mia-code/src/cli.ts)

- Flags: `--persona <name>`, `--list-personas`, `--persona-info <name>`

- Interactive persona selection mode for first-time users

5. Character Memory Integration (/src/mia-code/src/sessionStore.ts) - Extend storage schema: { sessionId, personaId, characterMemories[], narrativeState } - Implement CharacterMemoryManager for cross-session retrieval - Add semantic similarity search for relevant past interactions

Prototype Character Specification

Proposed First Character: "The Skeptical Cartographer"

• Core Function: Maps conceptual territories while questioning assumptions
• Voice Pattern: Tentative phrasing ("Perhaps...", "This suggests, though..."), frequent meta-commentary
• Memory Bias: Prioritizes contradictions and edge cases from previous sessions
• Interaction Style: Responds to certainty with doubt, to vagueness with structural questions

Academic Research Foundations

Research Question Formulation

Primary RQ: How do architectural patterns for agent persona management in dual-session LLM systems enable persistent narrative characterization that maintains coherence across context boundaries while supporting dynamic character development?

Secondary RQs:

1. What memory architectures best support character continuity in stateless LLM interactions?
2. How does the separation of functional and interpretive agents affect narrative consistency compared to monolithic character implementations?
3. What role does prompt engineering versus external memory play in perceived character authenticity?

Theoretical Framework Integration

Computational Narratology Foundations

The field examines algorithmic processes for creating and interpreting narratives through formal, computable representations. Your architecture embodies a narratological structure: the primary agent generates the fabula (event sequence), while the Unifier constructs the sjuzhet (narrative discourse).²³⁴

Recent work on multi-agent narrative systems demonstrates that LLMs can simulate nuanced characters when backed by memory of past interactions. Character consistency emerges from three architectural components: role-playing with memory, narrative coherence constraints, and persona-aware rewriting.⁵

Memory Architecture Patterns

Current research identifies critical memory tiers for LLM agents:⁶⁷

• Core memory (in-context, analogous to RAM): Character defining traits, immediate conversation
• Archival memory (vector store, analogous to disk): Past interaction embeddings, character experiences
• Entity memory: Structured capture of relationships, places, concepts tied to character knowledge

Your sessionStore.ts provides foundation for this hierarchy. MemGPT's approach of treating context as constrained resource with explicit memory management functions offers a proven pattern.⁷¹

Persona Consistency Research

Studies on personality consistency in conversational agents reveal that static predefined personas create "out-of-predefined persona" (OOP) problems when agents encounter queries beyond their initial description. Solutions involve:⁸⁹

• Dynamic persona retrieval from global collections based on dialogue context
• Natural Language Inference (NLI) models to ensure new persona elements align with core traits
• Posterior-scored architectures that weight persona relevance during generation

Your dual-session architecture naturally supports this: the Unifier can dynamically adjust interpretive framing based on retrieved character memories without altering the primary agent's functional execution.¹

Narrative Theory Convergence

Agent-to-Agent Communication as Narrative Protocol

Your interest in A2A protocols and Narrative Context Protocol (NCP) aligns with emerging work on computational narrative understanding. Recent frameworks model narrative as information flow from narrator to reader, where reader uncertainty and story model evolution become computable.¹⁰¹¹¹²

The Unifier functions as a computational narrator: it receives event summaries (fabula) and constructs character-filtered discourse (sjuzhet). This maps directly to narrative communication models where narration mediates between story world and audience.³

Indigenous Epistemology and Two-Perspective Architecture

Your Etuaptmumk (Two-Eyed Seeing) integration finds technical expression in the Mia/Miette duality. This ceremonial dual-perspective approach resonates with narrative theory on multiple focalization: the primary agent provides external focalization (events), while character-specific Unifiers offer internal focalization (interpretation).³¹

Computational approaches to narratives with non-human narrators explore how abstraction enables defamiliarization of anthropocentric assumptions. Your architecture's separation of "doer" and "reflector" creates space for non-Western narrative epistemologies to inform agent characterization.³

Literature Gaps and Contributions

Identified Research Gaps:

1. Most persona-based agents use monolithic architectures; dual-session approaches remain underexplored⁵⁸
2. Character memory systems focus on factual retrieval; narrative state and emotional continuity receive less attention¹³⁶
3. Computational narratology emphasizes plot structure over character interiority¹⁴¹⁵²
4. Agent personality research centers on chat consistency; character development over time is minimal¹⁶¹⁷

Your System's Novel Contributions:

• Architectural separation of functional cognition from narrative interpretation enables character layering without compromising task performance
• Session-aware memory infrastructure provides foundation for character development narratives across interactions
• Ceremonial framing (Miawa Unifier concept) brings indigenous epistemology into computational narrative generation
• CLI-based character selection democratizes access to narrative AI without GUI complexity

Article Structure Proposal

Title Options

1. "Dual-Session Architecture for Persistent Character Simulation in LLM-Based CLI Agents"
2. "From Functional Persona to Narrative Character: Memory and Architecture in Multi-Agent Systems"
3. "Ceremonial Computation: Two-Eyed Seeing in Agent Character Design"

Proposed Sections

I. Introduction: The Character Problem in Stateless Systems

• Context: LLMs as stateless transformers, the "conversational amnesia" challenge⁶
• Problem: Static personas vs. dynamic characters with development arcs
• Contribution: Dual-session architecture with persistent character memory

II. Related Work

• Computational narratology and formal narrative models¹¹⁴¹⁰
• Persona consistency in conversational agents⁹¹⁷⁸
• Memory architectures for LLM agents¹³⁷⁶
• Multi-agent narrative systems²¹⁴⁵

III. Architectural Foundations: The Unifier Model

• Dual-session flow: primary agent (Gemini) + interpretive agent (Claude)
• System prompt as character specification
• Session store as narrative memory substrate
• Comparison to monolithic persona approaches

IV. Character Memory Design

• Three-tier memory hierarchy (core/archival/entity)⁷⁶
• Character state schema: traits, experiences, relationships, narrative arc position
• Retrieval strategies for coherence: semantic similarity, temporal proximity, emotional resonance
• Cross-session continuity mechanisms

V. Implementation: The Skeptical Cartographer

• Prototype character specification
• Prompt engineering for consistent voice
• Memory integration examples
• User interaction flows

VI. Evaluation Framework

• Character consistency metrics: voice stability, trait adherence
• Narrative coherence: information flow accuracy, story model evolution¹⁰¹¹
• User perception: believability, engagement, emotional connection
• Performance: latency impact of memory retrieval, token efficiency

VII. Indigenous Epistemology in Agent Design

• Etuaptmumk as architectural principle (Two-Eyed Seeing = dual-session)
• Ceremonial framing vs. purely functional interaction
• Non-Western narrative structures in computational systems
• Ethical considerations in character simulation

VIII. Discussion

• Narrative agency vs. functional task completion
• Character development across sessions: growth, forgetting, contradiction
• Implications for human-AI collaboration in creative work
• Limitations: prompt brittleness, context window constraints, computational cost

IX. Future Directions

• Multi-character systems with inter-agent narrative dynamics
• Long-term memory consolidation strategies
• Integration with voice synthesis for embodied characters
• Applications in music therapy, ceremonial technology

Research Methods Outline

Technical Development Track:

1. Implement persona registry and dynamic prompt loading
2. Extend sessionStore with character memory schema
3. Build prototype character (Skeptical Cartographer)
4. Develop CLI persona selection interface
5. Create character memory query/update API

Empirical Evaluation Track:

1. Consistency Testing: Generate 100+ interactions per character across varied queries, measure voice/trait stability using NLI similarity scores⁸
2. User Study: 20-30 participants interact with 3 characters over 5 sessions each, gather qualitative feedback on believability and quantitative engagement metrics
3. Memory Effectiveness: Ablation study comparing no-memory vs. session-only vs. cross-session memory on coherence scores¹⁰
4. Performance Benchmarking: Measure latency and token usage across memory configurations

Theoretical Analysis Track:

1. Map system to formal narrative models (narrator-reader information flow)¹¹¹⁰
2. Analyze through indigenous epistemology lens (interviews with Etuaptmumk practitioners)
3. Compare to existing computational narratology frameworks¹²⁴

GitHub Development Roadmap

Phase 1: Foundation (Weeks 1-3)

• Create /src/mia-code/src/personas/ directory structure
• Define IPersona TypeScript interface
• Implement PersonaRegistry class
• Add configuration schema for activePersona

Phase 2: Character Infrastructure (Weeks 4-6)

• Build CharacterMemoryManager extending sessionStore
• Implement three-tier memory (core/archival/entity)
• Create retrieval algorithms: semantic search, temporal weighting
• Add persistence layer for cross-session character state

Phase 3: Prototype Character (Weeks 7-9)

• Author Skeptical Cartographer system prompt
• Define character memory schema (traits, biases, experiences)
• Implement character-specific memory update logic
• Create test suite for consistency validation

Phase 4: CLI Integration (Weeks 10-11)

• Add --persona flag parsing
• Implement --list-personas display
• Create interactive persona selection menu
• Add character info display command

Phase 5: Evaluation Tools (Weeks 12-14)

• Build consistency testing harness
• Implement coherence metrics (NLI-based)⁸
• Create user study instrumentation
• Develop performance profiling tools

Phase 6: Documentation & Examples (Weeks 15-16)

• Update rispecs for character system
• Create character authoring guide
• Write example characters (3-5 distinct personas)
• Generate demo interaction transcripts

Citation Infrastructure

Key Papers to Cite:

• Computational narratology foundations: Mani (2013), Finlayson (2021)⁴¹¹¹⁰
• Narrative generation with LLMs: Yu (2025), Ammanabrolu (2020)¹⁸⁵
• Agent personality consistency: Liu (2022)⁹⁸
• Memory architectures: MemGPT/Letta documentation, Serokell LLM memory patterns⁶⁷
• Multi-agent narratology: Balestri (2025)¹⁹¹⁴
• AI narrative ethics: Caracciolo (2025)³

Datasets for Evaluation:

• ConvAI2 (persona-based dialogue)⁹⁸
• ROCStories (narrative coherence benchmarks)
• Custom CLI interaction corpus (to be collected)

Theoretical Foundations:

• Narratology: Genette (focalization), Bal (fabula/sjuzhet)
• Indigenous epistemology: Marshall & Bartlett (Etuaptmumk/Two-Eyed Seeing)
• Computational narrative: Prince, Herman (cognitive narratology)

---

This framework synthesizes your codebase investigation with current research to provide both immediate development steps and longer-term academic positioning. The dual-session architecture you've built is well-aligned with emerging work on agent memory and computational narratology, while your indigenous epistemology integration offers a genuinely novel contribution to the field.

The software pathway is concrete and modular: persona prompt library → config extension → dynamic selection → memory integration → CLI exposure. Each phase builds on existing infrastructure without requiring architectural rewrites.

The academic pathway positions this work at the intersection of computational narratology, agent memory systems, and indigenous knowledge integration—a space with clear research gaps and strong theoretical foundations. The proposed evaluation combines technical metrics (consistency, coherence) with human factors (believability, engagement) and cultural dimensions (ceremonial framing, non-Western narrative structures). <span style="display:none">^{202122232425262728293031323334}</span>

<div align="center">⁂</div>

Footnotes

1. persona-to-narrative-character-inquiry_260201.md ↩ ↩² ↩³ ↩⁴

2. https://arxiv.org/abs/2509.00481 ↩ ↩² ↩³

3. https://muse.jhu.edu/article/980959 ↩ ↩² ↩³ ↩⁴ ↩⁵

4. https://www-archiv.fdm.uni-hamburg.de/lhn/node/43.html ↩ ↩² ↩³ ↩⁴

5. https://aclanthology.org/2025.in2writing-1.9.pdf ↩ ↩² ↩³ ↩⁴

6. https://serokell.io/blog/design-patterns-for-long-term-memory-in-llm-powered-architectures ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

7. https://www.letta.com/blog/agent-memory ↩ ↩² ↩³ ↩⁴ ↩⁵

8. https://arxiv.org/abs/2208.10816 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

9. https://openreview.net/pdf?id=90KaW3XjgRq ↩ ↩² ↩³ ↩⁴

10. https://arxiv.org/pdf/2103.12872.pdf ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

11. https://www.aclweb.org/anthology/2021.nuse-1.10 ↩ ↩² ↩³ ↩⁴ ↩⁵

12. https://aclanthology.org/2021.emnlp-main.26.pdf ↩ ↩²

13. https://www.emergentmind.com/topics/persistent-memory-for-llm-agents ↩ ↩²

14. https://www.scitepress.org/DigitalLibrary/Link.aspx?doi=10.5220/0013369600003890 ↩ ↩² ↩³

15. https://arxiv.org/abs/2508.07010 ↩

16. https://www.sciencedirect.com/science/article/pii/S0747563225002067 ↩

17. https://cetas.turing.ac.uk/publications/patterns-not-people-personality-structures-llm-powered-persona-agents ↩ ↩²

18. https://www.aclweb.org/anthology/2020.conll-1.42.pdf ↩

19. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5165274 ↩

20. https://data.doi.or.kr/10.23362/KOEN2025.07.25.1.031 ↩

21. https://www.mdpi.com/1424-8220/25/17/5265 ↩

22. http://services.igi-global.com/resolvedoi/resolve.aspx?doi=10.4018/978-1-5225-0432-0.ch001 ↩

23. https://www.semanticscholar.org/paper/c750fb0623c1a4c9cd53c59e0d2381cef5278b45 ↩

24. http://biorxiv.org/lookup/doi/10.1101/2024.06.04.595972 ↩

25. https://www.mdpi.com/2075-4426/15/7/298 ↩

26. https://arxiv.org/html/2503.23512v1 ↩

27. https://arxiv.org/pdf/2503.04844.pdf ↩

28. https://www.aclweb.org/anthology/W18-6602.pdf ↩

29. https://arxiv.org/html/2405.13042v2 ↩

30. https://www.cambridge.org/core/journals/computational-humanities-research/announcements/call-for-papers/computational-narratology ↩

31. https://www.scitepress.org/publishedPapers/2025/133696/pdf/index.html ↩

32. https://journal.accsindia.org/show.article.php?id=59 ↩

33. https://www.uib.no/sites/w3.uib.no/files/attachments/ai-stories-b1-erc_adg.pdf ↩

34. https://www.sciencedirect.com/science/article/abs/pii/S1875952125001570 ↩