Smriti: Version Control for Reasoning

Smriti is an innovative tool designed to enhance reasoning processes by providing version control specifically tailored for conversations involving large language models (LLMs) and agentic reasoning.

Problem Statement

The project was conceived from the frustration of continually losing track of reasoning states while switching between different AI models like ChatGPT and Claude. Frequent context switching often led to the necessity of reconstructing thoughts and conclusions from scratch, which can be time-consuming and inefficient. Smriti aims to eliminate this bottleneck by enabling users to easily manage and return to their reasoning states.

Core Features

• Checkpoint Creation: Capture structured snapshots of reasoning progress that include decisions made, objectives, open questions, and tasks that need to be accomplished. This allows users to revisit any prior state and mount checkpoints to continue their thought process seamlessly.

• Model Switching: Users can switch AI models mid-session without the need to rewrite extensive context each time, enhancing productivity.

• Branching and Comparing: Smriti supports branching reasoning where users can fork their thought processes and experiment with different directions without losing the original context. Checkpoints can also be compared to visualize reasoning divergences.

Unique Approach

Unlike typical chat systems that rely on ambient prompts, Smriti enforces strict boundaries within its data layer to regulate contextual visibility. This ensures that only relevant turns are displayed without leaks from future or unrelated sessions. This distinct methodology aims to enhance clarity and focus during reasoning sessions.

Applications Beyond Chat:

While initially intended for conversational use, Smriti offers significant advantages for agents that engage in multi-step reasoning. The tool provides a structured way to debug processes, reproduce results, and analyze outcomes effectively.

User Engagement

For a better understanding of Smriti's capabilities, a demonstration is available here. The visual representation of features will assist in grasping how checkpoints function and how they can influence the reasoning flow.

Limitations

As of now, Smriti operates as a single-user application and does not support branch merging or streaming responses. The interface is functional but not polished, and there are no mobile integrations yet.

Technical Stack

Smriti is built using FastAPI for the backend and React with TypeScript for the frontend, utilizing PostgreSQL for data storage. This combination ensures robust performance and scalability for future development.

Future Prospects

While the current version focuses on validating concepts, there are numerous potential directions for enhancement, such as structured reasoning memory and shared states across various tools. Feedback on the approach is welcome to guide future development efforts.

Smriti represents a thoughtful progression toward streamlined reasoning practices in an era where efficient thought management is increasingly crucial.