Mstar

LLM agents need memory to store what they have seen and retrieve it when needed. Everyone hand-designs these systems: pick a vector store, write some retrieval logic, tune the prompts, ship it. This project automates the design process. You provide a benchmark; it evolves a task-specific memory system as executable Python code. The memory it discovers for conversational QA looks nothing like what it discovers for a household robot. Both emerge from the same three simple seeds.

The idea

A memory system is expressed as a *memory program* — a Python module with three jointly-optimized dimensions: Schema (dataclasses defining what to store and how to query), Logic (`write()`/`read()` backend), and Instruction (prompt constants that steer the task agent). In the code, this is the `KBProgram` (defining `KnowledgeItem`, `Query`, and `KnowledgeBase`); in the paper it is called the *memory program* or *memory harness* — they are the same thing. The evolution loop starts from 3 seeds, evaluates programs on a benchmark (scoring against a static validation set and using a rotating set for