Give one OpenClaw session durable memory, topic-aware continuity, and bounded token growth.

Given a bug and a token budget, how should you pick which code to show the LLM? This repo contains a benchmark that evaluates different context assembly methods on that question, using controlled single-line mutations across two Python repositories. Still in development, will expand to more methods, models and repositories.