W-GRAM-LM: A Research Framework for World-Guided Recursive Attractor Language Models

W-GRAM-LM is an open-source research codebase focused on world-guided recursive language modeling. It integrates cutting-edge technologies including latent world prediction, multi-trajectory reasoning, and answer attractor convergence, supporting research on auditable agent memory and reasoning architectures. The project is maintained by edu-ide, with source code hosted on GitHub (https://github.com/edu-ide/wgram-lm), licensed under AGPL-3.0, and released on May 31, 2026.

Current agent memory and reasoning systems face auditability challenges: components like implementation code and training workflows are scattered across private systems, making it difficult for researchers to understand the interactions between various parts. The mission of W-GRAM-LM is to integrate these components into a unified codebase to support reproducible research.

Latent World Prediction

Integrates LeWorldModel and JEPA-style latent state prediction probes to capture abstract world dynamics, supporting multi-step reasoning and causal understanding.

Multi-Trajectory Reasoning

Adopts GRAM/PTRM style to explore multiple reasoning paths simultaneously, increasing diversity through random recursive breadth.

Answer Attractor

Guides the model to converge to consistent answers through attractor dynamics while maintaining generation diversity.

Stable Recursion

Parcae-style mechanisms ensure recursive numerical stability and long-term memory retention.

MemoryOS Memory System

Integrates LLM Wiki, Harrier text embeddings, and visual embeddings to provide multi-modal context support.

Standalone Smoke Test Mode

No need to download Qwen weights; trains and tests small random models to lower the experimental threshold.

Donor Adapter Mode

Loads Qwen3.5-style models, freezes most weights, and only trains modules like adapters to reduce resource requirements.

1. Reproducible experimental environment for predictive language models;
2. Supports ablation studies of components like random recursive breadth and answer attractors;
3. Small-scale smoke tests are suitable for resource-constrained scenarios;
4. Auditable multi-modal retrieval pipeline;
5. Documents architectural decisions to pass on knowledge.

Uses the AGPL-3.0 strong copyleft license, requiring derivative works to be open-sourced. Includes governance documents such as CODE_OF_CONDUCT and CONTRIBUTING to ensure the healthy development of the community.

W-GRAM-LM integrates cutting-edge technologies to provide an auditable and reproducible open-source platform for agent memory and reasoning research. Its core value lies in helping researchers clearly understand the internal mechanisms of the system, making it a project worthy of attention for researchers in related fields.