BACM: Transparent, Provenance-Aware Shared Memory for Parallel LLM Agents on Changing Codebases

When several LLM agents work on the same codebase in parallel, each re-reads the same files and re-derives the same facts, and once the code changes, every agent keeps relying on what it read before. Shared memory removes the redundancy but makes staleness worse: a stale fact misleads every agent that reuses it. We present BACM (Branch-Aware Context Manager), a shared context manager that stores knowledge as typed objects with provenance (file, version, supporting span), dependency edges and derivation rules. When a source changes, BACM invalidates exactly the affected part of the dependency graph and recomputes only what an open question needs. A first design that exposed the memory to the model as tools cost more than independent agents, so BACM is transparent: the tool layer publishes what agents read, serves already-read files as digests and answers questions from known facts by typed rules. We evaluate BACM on RepoQA-Dyn, a benchmark of five real Python repositories in which a real release lands mid-run, with ground truth from static analysis. In a pre-registered evaluation on held-out seeds with GPT-5 mini, BACM answered 89% of questions correctly on static repositories and 88% when the repository changed, against 58% and 49% for independent agents with context eviction (both significant), and cut stale answers after the release from 61 to 7 of 186. Reasoning quality is not state consistency: agents that reason more are the most often stale. Token savings (-22% and -26%) are not statistically significant and vary by repository. Two implementation defects found after the pre-registered runs are reported with frozen and corrected results. Code: https://github.com/tejashvi-kumawat/bacm