wtf is yams?

YAMS, short for Yet Another Memory System, is a research project that studies when giving a language model access to an external memory actually improves its answers, and when it makes no difference at all. The author's central finding is that memory only matters when there is a gap between what a model can work out or check for itself from the evidence in front of it, and what the memory claims. When that gap is open, a correct memory helps and a wrong one hurts. When the model could already verify the answer on its own, the memory has no effect either way. The repository backs this claim with experiments run across three model sizes, comparing answers with no memory against answers using a validated rule stored in memory. Smaller models showed a large accuracy gain from good memory, a mid-sized model showed a smaller and less certain gain, and the largest model performed identically with or without memory because it could already verify the answers itself. A separate test showed the same single model rejecting a wrong memory when it could check it, but being misled by that same wrong memory when it could not, showing the effect comes from whether an answer is checkable rather than from which model is used. The full technical report is included as a PDF, and the repository provides several small, dependency-free Python scripts in a benchmarks folder that regenerate the test questions and reproduce the core results. These require Python 3.10 or later and nothing else to run. The heavier system that produced some of the original results, built with FastAPI and SQLite, is not included, but its raw output data is provided separately for reference. There is also a working example of the memory system in use: an MCP server in the mcp folder that plugs a validated memory into a coding agent, letting the agent recall stored rules and check its own actions against them. The author is upfront that this is a single person's exploratory study with a small number of trials, not a finalized or peer-reviewed result, and lists the study's limitations directly in the README.