mayiwen0212/memchain

MemChain is a research project that gives AI chat systems a more transparent way of deciding which pieces of conversation history to remember and use when answering a question. Instead of quietly stuffing whatever text a search step retrieves into the AI's context, MemChain produces an explicit, readable trace of its memory decisions before generating an answer. Given a new question and a pool of candidate memories drawn from earlier conversation, the system first figures out what kind of evidence the question actually needs. It then decides, memory by memory, whether to keep it, drop it, refine it, merge it with another, or stop looking further, and it records this as an ordered chain of steps that cite which memories were used at each point. Only a final, compact set of these approved memories is passed along to a separate model that generates the actual answer, keeping the memory selection process separate from and easier to check than the answer itself. The candidate memory pool itself is built only from the question and the conversation's own history, never from the correct answer, which keeps the process honest and testable rather than quietly shaped by hindsight. The repository includes a small runnable example with sample dialogue sessions and questions so you can see the full output format, including the intent analysis, the memory actions taken, the chain of cited steps, and the final memories passed to the answer model. The full training setup described in the paper uses eight high end GPUs and an approach that first fine tunes the system on examples, then further trains it with reinforcement learning, though the private training data, model checkpoints, API keys, and machine specific scripts used for that training are not included here. What is included is the core method code: the memory pool construction logic, the schema definitions, prompt templates, reward calculations, and a small test suite, along with optional support for open source embedding models if you want dense semantic search over memories rather than plain keyword matching. This is an academic code release intended for inspection and extension by other researchers, and it is licensed under the MIT license.