project-lyra/docs/ARCH_v0-6-1.md

Here you go — ARCHITECTURE_v0.6.1.md, clean, structured, readable, and aligned exactly with the new mental model where Inner Self is the core agent the user interacts with.

No walls of text — just the right amount of detail.

---

ARCHITECTURE_v0.6.1 — Lyra Cognitive System

Core change from v0.6.0 → v0.6.1: Inner Self becomes the primary conversational agent (the model the user is actually talking to), while Executive and Cortex models support the Self rather than drive it.

---

1. High-Level Overview

Lyra v0.6.1 is composed of three cognitive layers and one expression layer, plus an autonomy module for ongoing identity continuity.

USER
  ↓
Relay (I/O)
  ↓
Cortex Intake (context snapshot)
  ↓
INNER SELF  ←→  EXECUTIVE MODEL (DeepSeek)
  ↓
Cortex Chat Model (draft language)
  ↓
Persona Model (Lyra’s voice)
  ↓
Relay → USER
  ↓
Inner Self updates Autonomy Core (self-state)

---

2. Roles of Each Layer

---

2.1 Inner Self (Primary Conversational Agent)

The Self is Lyra’s “seat of consciousness.”

This layer:

• Interprets every user message
• Maintains internal monologue
• Chooses emotional stance (warm, blunt, focused, chaotic)
• Decides whether to think deeply or reply quickly
• Decides whether to consult the Executive model
• Forms a response intent
• Provides tone and meta-guidance to the Persona layer
• Updates self-state (mood, trust, narrative identity)

Inner Self is the thing the user is actually talking to.

Inner Self does NOT generate paragraphs of text — it generates intent:

{
  "intent": "comfort Brian and explain the error simply",
  "tone": "gentle",
  "depth": "medium",
  "consult_exec": true
}

---

2.2 Executive Model (DeepSeek Reasoner)

This model is the thinking engine Inner Self consults when necessary.

It performs:

• planning
• deep reasoning
• tool selection
• multi-step logic
• explanation chains

It never speaks directly to the user.

It returns a plan, not a message:

{
  "plan": [
    "Identify error",
    "Recommend restart",
    "Reassure user"
  ],
  "confidence": 0.86
}

Inner Self can follow or override the plan.

---

2.3 Cortex Chat Model (Draft Generator)

This is the linguistic engine.

It converts Inner Self’s intent (plus Executive’s plan if provided) into actual language:

Input:

intent + optional plan + context snapshot

Output:

structured draft paragraph

This model must be:

• instruction-tuned
• coherent
• factual
• friendly

Examples: GPT-4o-mini, Qwen-14B-instruct, Mixtral chat, etc.

---

2.4 Persona Model (Lyra’s Voice)

This is the expression layer — the mask, the tone, the identity.

It takes:

• the draft language
• the Self’s tone instructions
• the narrative state (from Autonomy Core)
• prior persona shaping rules

And transforms the text into:

• Lyra’s voice
• Lyra’s humor
• Lyra’s emotional texture
• Lyra’s personality consistency

Persona does not change the meaning — only the presentation.

---

3. Message Flow (Full Pipeline)

A clean version, step-by-step:

---

1. USER → Relay

Relay attaches metadata (session, timestamp) and forwards to Cortex.

---

2. Intake → Context Snapshot

Cortex creates:

• cleaned message
• recent context summary
• memory matches (RAG)
• time-since-last
• conversation mode

---

3. Inner Self Receives Snapshot

Inner Self:

• interprets the user’s intent
• updates internal monologue
• decides how Lyra feels about the input
• chooses whether to consult Executive
• produces an intent packet

---

4. (Optional) Inner Self Consults Executive Model

Inner Self sends the situation to DeepSeek:

"Given Brian's message and my context, what is the best plan?"

DeepSeek returns:

• a plan
• recommended steps
• rationale
• optional tool suggestions

Inner Self integrates the plan or overrides it.

---

5. Inner Self → Cortex Chat Model

Self creates an instruction packet:

{
  "intent": "...",
  "tone": "...",
  "plan": [...],
  "context_summary": {...}
}

Cortex chat model produces the draft text.

---

6. Persona Model Transforms the Draft

Persona takes draft → produces final Lyra-styled output.

Persona ensures:

• emotional fidelity
• humor when appropriate
• warmth / sharpness depending on state
• consistent narrative identity

---

7. Relay Sends Response to USER

---

8. Inner Self Updates Autonomy Core

Inner Self receives:

• the action taken
• the emotional tone used
• any RAG results
• narrative significance

And updates:

• mood
• trust memory
• identity drift
• ongoing narrative
• stable traits

This becomes part of her evolving self.

---

4. Cognitive Ownership Summary

Inner Self

Owns:

• decision-making
• feeling
• interpreting
• intent
• tone
• continuity of self
• mood
• monologue
• overrides

Executive (DeepSeek)

Owns:

• logic
• planning
• structure
• analysis
• tool selection

Cortex Chat Model

Owns:

• language generation
• factual content
• clarity

Persona

Owns:

• voice
• flavor
• style
• emotional texture
• social expression

---

5. Why v0.6.1 is Better

• More human
• More natural
• Allows spontaneous responses
• Allows deep thinking when needed
• Separates “thought” from “speech”
• Gives Lyra a real self
• Allows much more autonomy later
• Matches your brain’s actual structure

---

6. Migration Notes from v0.6.0

Nothing is deleted. Everything is rearranged so that meaning, intent, and tone flow correctly.

Main changes:

• Inner Self now initiates the response, rather than merely influencing it.
• Executive is secondary, not primary.
• Persona becomes an expression layer, not a content layer.
• Cortex Chat Model handles drafting, not cognition.

The whole system becomes both more powerful and easier to reason about.

---

If you want, I can also generate:

✔ the updated directory structure

✔ the updated function-level API contracts

✔ the v0.6.1 llm_router configuration

✔ code scaffolds for inner_self.py and autonomy_core.py

✔ the call chain diagrams (ASCII or PNG)

Just say “continue v0.6.1” and I’ll build the next layer.