We present Memor, a Python library for structuring, managing, and transferring conversation histories across different large language models through an object-oriented, Pythonic interface. Visitors learn how the library abstracts LLM interactions into intuitive Session objects that encapsulate sequences of message exchanges, including not only the content but also critical metadata such as decoding temperature, token counts, and model-specific parameters, enabling comprehensive and reproducible logs of interactions.
The poster demonstrates the technical architecture, including the session management system, message filtering capabilities, and cross-model transfer mechanisms, while showcasing practical examples that illustrate how users can seamlessly migrate conversation context between different LLMs. An example of this could be starting the conversation with a Retrieval Augmented Generation (RAG) model to gather relevant information and then switching to a reasoning-specialized model to solve problems based on the retrieved context.
Through code demonstrations, we show how Memor's structured data format enables users to select, filter, and share specific portions of past conversations. These features facilitate granular control over which messages and context are preserved or transferred. The presentation discusses practical applications from multi-model workflows where different LLMs are leveraged for their specialized capabilities to research applications requiring reproducible LLM experiments with complete parameter tracking, and collaborative scenarios where conversation histories need to be shared or archived in a scientific report for reproducibility.
We demonstrate Memor’s intuitive API, its flexibility across LLM providers and message formats, and its extensibility for custom prompt templates. We also show how Memor simplifies conversation management in increasingly complex LLM applications by standardizing context handling across models. This positions Memor as both a practical tool for developers and a framework for improving research reproducibility in NLP.