Dynamic Memory First Collaborative Agent Long-Form Creativity

CoLong Idea Studio

A Dynamic-Memory-First Collaborative Agent Framework for Long-Form Creative Ideation and Story Generation

Independent Research 2026
A research-facing system for collaborative idea refinement, chapter planning, dynamic memory retrieval, and controllable long-form generation.
Agentic
Collaborative ideation loop before writing
Memory-First
Outlines, facts, world settings, and chapter summaries
Observable
Progress log exposes hidden planning and memory signals

Overview

CoLong Idea Studio targets long-form, chaptered, and high-consistency creative writing tasks. The framework is organized around a dynamic-memory-first generation loop, where planning artifacts, chapter summaries, character settings, world settings, and fact cards are continuously written back and later retrieved as contextual constraints.

In addition to generation itself, the project emphasizes collaborative ideation. Before the drafting phase begins, the system engages the user in an iterative clarification process so that vague premises can be transformed into a more stable creative brief. This improves downstream coherence, reduces prompt underspecification, and better aligns the system with human-centered creative workflows.

Project Links

Current focus: collaborative ideation, dynamic-memory-guided story generation, progress-log observability, and deployment-ready packaging.

Core Contributions

Collaborative Ideation Agent

The idea-copilot procedure is implemented as a genuine agent loop that keeps asking targeted questions until the user explicitly confirms the concept is ready.

Dynamic Memory Priority

The system prioritizes dynamic memory over static retrieval, storing and recalling chapter text, outlines, fact cards, character settings, and world settings.

Progress-Log Observability

Runtime logs expose global outline creation, chapter plans, chapter outlines, inferred length targets, memory snapshots, and setting writes.

Completion-First Chapter Execution

The primary objective is to complete planned chapters rather than stopping early when an external evaluator score happens to cross a threshold.

System Workflow

The workflow figure below summarizes the full trajectory from idea refinement to outline planning, chapter execution, dynamic-memory writeback, and subsequent contextual reinjection.

CoLong Idea Studio workflow diagram

Methodology

01

Outline-Grounded Length Steering

Chapter length is treated as a prompt-level target inferred from outline semantics, not as a simplistic hard cap detached from narrative structure.

02

Typed Dynamic Memory Assembly

Retrieved context is grouped by semantic role, including outlines, facts, characters, and world settings, before being injected back into generation prompts.

03

User-Confirmed Ideation Exit

The collaborative ideation loop terminates only when the user explicitly confirms readiness, making the transition into drafting both intentional and inspectable.

Dynamic Memory and Progress Log

Memory Buckets

  • texts
  • outlines
  • characters
  • world_settings
  • plot_points
  • fact_cards

These structured buckets are maintained in memory_index.json and serve as lightweight narrative anchors across chapters.

Representative Log Events

global_outline
chapter_outline_ready
chapter_plan
chapter_length_plan
character_setting
world_setting
memory_snapshot

The logging layer reveals hidden planning and memory signals, making long-form generation easier to debug and inspect.

Evaluation Snapshot

The current project page retains an evaluation-style visualization for research presentation while the main system focus has shifted toward collaborative ideation quality, generation observability, and dynamic-memory consistency.

Automatic evaluation snapshot

Deployment and Usage Notes

Runtime Entry Points

CLI: python main.py

Web Portal: python -m uvicorn local_web_portal.app.main:app --host 0.0.0.0 --port 8010

Portal Goal: support collaborative ideation, long-form planning, and chapter generation for multi-user deployment.

Recommended Packaging Principle

For server deployment, keep only runtime-required files and exclude historical outputs, local caches, transient vector data, and environment artifacts whenever possible.

This reduces repository noise, improves cold-start clarity, and better matches a reproducible research-demo workflow.

Star Growth

Open Full Chart

Repository momentum is often a more useful signal here than a citation block, so the project page now highlights GitHub star growth instead.