🌐 Language / 言語

• 🇺🇸 English (current)
• 🇯🇵 日本語版はこちら

AI conversations often collapse in predictable ways.

A structural model for stabilizing long AI conversations and creating focused reasoning environments.

Long conversations often collapse because all tokens compete for attention inside a single linear context.

As discussions grow, references drift, assumptions blur, and reasoning paths interfere with each other.

BRM introduces a scoped reference structure designed to organize conversation context, preserve reasoning continuity, and help AI remain focused on the current problem space.

If you came here from X or Zenn:

AI conversations often fail in predictable ways:

• context drift
• reasoning instability
• prompt breakdown over time

This is not random.

Start with a real example:

→ Case Study (see failure first)

Most AI workflows rely on prompt control:

Prompt
↓
Model
↓
Output

Prompts can strongly influence the initial response.

However, during longer conversations their influence often weakens.

This repository refers to this phenomenon as Prompt Dissolution — the gradual weakening of prompt influence as conversational context expands.

As conversational context grows, model behavior may increasingly follow:

• conversational coherence
• probabilistic completion
• internal reference structures

rather than the original prompt instructions.

The Branching Reference Model (BRM) explores a structural alternative.

Instead of relying solely on prompts, BRM investigates how reference structures may stabilize long interaction chains.

Prompt

• Structural Anchors
• Controlled Reference Scope
  ↓
  Stable Long-Term Collaboration

BRM organizes discussion into recoverable reference structures, allowing reasoning paths to remain distinct even as interaction history expands.

The goal is not simply to make AI remember more.

The goal is to create an environment where AI can:

• remain oriented
• focus on one topic or problem space at a time
• reason from an explicit reference position
• return to earlier reasoning paths without collapsing them into one linear context

BRM Core v3 moves the model from reference philosophy toward an implementable protocol.

→ Read BRM Core v3

v3 defines BRM as a governance layer built around a multi-layer semantic graph with cognitive coordinates.

Its central idea is simple:

AI should not need to reconstruct the entire history of a conversation every time it reasons.

Instead, BRM explicitly manages:

• where the current reasoning is located
• which Nodes and Series are relevant
• which references should be full, summarized, label-only, or outside scope
• when thought moves to a new Node or Series
• when something should remain unresolved rather than being forced into a conclusion

BRM v3 also makes several boundaries explicit:

Artifact ≠ Node

Generated documents, code, images, PDFs, spreadsheets, and other outputs may be referenced by Nodes, but the artifacts themselves are not cognitive Nodes.

File existence ≠ File reference permission

Files remain outside the active reference scope unless they are explicitly loaded.

Transition ≠ Closure

Moving to another Series or Hub does not require the previous one to be finished.

Unknown is a valid state

Undefined, unresolved, pending, or uncertain structures may remain open until later reasoning provides enough information.

This allows BRM to preserve the trajectory of thought without forcing premature organization or resolution.

If you'd like to experience one of these structures in practice:

→ Download Symptom Stable v1.2

If you want to understand why BRM exists:

→ Read the Origin Timeline

If you're new, start with:

→ Case Study (Start Here)

Then:

→ Stable Thinking Stack

Then read the current core model:

→ BRM Core v3

After that, explore freely.

👉 Understand where BRM fits among prompts, RAG, and agents
→ Concept Comparisons

Current AI systems often reason over linear conversational context.

As interactions grow longer, this structure creates several problems:

• references drift across unrelated parts of the conversation
• reasoning paths interfere with each other
• important assumptions become diluted by unrelated tokens
• older hypotheses, abandoned ideas, generated artifacts, and current decisions may remain inside the same reference stream

These effects often appear as:

• hallucinations
• inconsistent reasoning
• sudden loss of context
• unstable conclusions
• confusion about what is currently authoritative

However, many of these failures may be structural rather than purely model limitations.

BRM proposes that stabilizing the reference structure of conversations can improve long-term collaboration with AI systems.

Instead of treating conversation history as a single expanding stream, BRM introduces structured reference organization.

This allows reasoning paths to remain distinct, recoverable, and selectively referenceable even as conversations grow large.

The Branching Reference Model separates three conceptual layers of AI collaboration:

Conversation Structure
↓
Reasoning Process
↓
Inference Engine

In this view:

User interacts with the language model
↓
The conversation context is organized using BRM reference structures
↓
Stable Modes may guide reasoning behavior
↓
The language model performs inference

BRM organizes the structure of discussion and reference.

Stable Modes guide reasoning posture.

The language model performs inference.

These layers are intentionally separable.

Changing a reasoning mode does not necessarily change the current Node or Series.

Likewise, moving through the BRM reference structure does not require changing the reasoning mode.

This separation allows reasoning strategies to evolve without destabilizing conversation structure.

BRM treats thought as a multi-layer semantic graph with cognitive coordinates.

A simplified view:

Hub
↓
Series
↓
Node Chain

Nodes preserve observable cognitive transitions.

Series preserve larger thematic reasoning arcs.

Hubs group broader semantic regions.

Link Edges connect related cognitive events across the structure.

Reference Scope determines what the AI may actively use during the current reasoning step.

The purpose is not to load everything.

The purpose is to provide the right reference space for the current reasoning task.

Reference Scope is one of the central mechanisms in BRM Core v3.

A BRM implementation may distinguish references such as:

Active Reference
Full Reference
Summary Reference
Label Reference
Outside Scope

This means stored information and currently usable information are not treated as the same thing.

Something may exist in the BRM structure without being included in the current AI context.

This is particularly important for:

• older reasoning branches
• unrelated Series
• generated artifacts
• uploaded files
• unresolved material
• historical decisions that are not currently relevant

BRM therefore treats reference selection as part of reasoning governance.

BRM does not require every thought structure to be immediately resolved.

A Node, Series, Hub, Link, Label, or routing decision may remain:

• open
• pending
• unresolved
• undefined
• unknown

Moving elsewhere does not automatically close the previous structure.

This is intentional.

Human thought often moves before it finishes.

BRM preserves that behavior rather than forcing every transition into a completed state.

BRM does not require a multi-agent system.

However, one possible implementation separates two roles:

Main AI
↓
answers and reasons inside the current Reference Scope

Observer AI
↓
observes structural movement

The Observer may detect:

• Node transitions
• Series boundaries
• Link candidates
• RETURN / NEW routing
• unresolved states

A BRM engine can then perform the mechanical operations:

• Node storage
• coordinate management
• Link creation
• Reference Scope construction
• navigation label management

Conceptually:

The Main AI thinks.
The Observer records where the thinking moved.

This can reduce the amount of structural management required from the answering model itself.

This repository explores structural failures in long AI collaboration and proposes a model to address them.

Conceptual progression:

Observed collaboration failures
↓
Prompt dissolution
↓
Reference instability
↓
Stable reasoning environments
↓
Branching Reference Model

The repository contains two main parts:

• Conceptual essays describing observed structural issues in AI conversations
• The BRM core model describing the proposed structural framework

These essays explore structural issues observed in long AI conversations.

→ Perspective differences between users and AI during reasoning
→ Read essay

→ How reference drift appears in long AI conversations
→ Read essay

→ The idea of stabilizing reasoning through structured reference anchors
→ Read essay

→ The hidden failure mode of RAG and enterprise AI
→ Why “feeding everything into AI” does not create veteran judgment
→ Read essay

→ Why role-specific AI needs reference boundaries, not just prompts
→ How role-based AI can multiply contamination when every role shares the same archive
→ Read essay

→ Why persistent agents need branch-aware reference architecture
→ How long-lived AI can preserve memory without turning artifacts, failures, and reactions into false continuity
→ Read essay

→ Why structure may matter as much as model capability
→ Read

The current core specification:

→ BRM Core v3

Previous version:

→ BRM Core v2

BRM Core v3 focuses on making the reference philosophy implementable.

It introduces or formalizes:

• multi-layer semantic structure
• cognitive coordinates
• explicit Reference Scope
• semi-automatic Node / Series transitions
• open and unresolved states
• immutable cognitive history
• Link-based reinterpretation
• artifact / Node separation
• explicit file reference isolation
• optional Main AI / Observer AI separation
• reasoning posture as a layer independent from structural navigation

→ A practical entry point for structured reasoning
→ Download Symptom Stable v1.2

If you want to explore deeper:

• → Concept Comparisons
  Understand how BRM differs from prompts, RAG, and agents

• → Case Studies
  See real examples of failure and structural recovery

• → Stable Thinking Stack
  Apply structured reasoning in practice

• → BRM Core v3
  Explore the current formal structure and implementation model

• → BRM Core v2
  Read the previous core model

• → Architecture Before Engine
  Understand the broader design philosophy

Readers can explore freely after following the main path:

1. Case Study
2. Stable Thinking Stack
3. BRM Core v3
4. Conceptual Exploration

From the system side, BRM is a reference-governance protocol.

From the user side, the experience may feel much simpler:

an expandable thought space where ideas can branch, pause, remain unresolved, reconnect, and return — while the AI does not lose the reference point.

A question can become another question.

An unfinished idea can remain unfinished.

An older path can be revisited from a different angle.

Multiple reasoning spaces can remain open at the same time.

The purpose is not to force human thought into a rigid tree.

The purpose is to let thought expand while preserving enough structure for both the user and AI to know:

where they are, what they are referring to, and where they can return.

In that sense, BRM aims for a simple experience:

Think as widely as you want.
The AI should not get lost.