Long-horizon conversational agents require memory that can evolve with dialogue, retrieve information on demand, and remain uncertainty-aware when evidence is insufficient, rather than merely larger storage or context windows. Existing methods store interaction histories as retrievable text, summaries, or vector indices, and retrieve relevant content at query time. Although effective for simple factual queries, this paradigm does not explicitly model how memory writes change state, leaving temporal updates, plan changes, and counterfactual reasoning to query-time inference over retrieved text.

We propose ProgMem, which compiles dialogue events into structured memory programs and checks answers against dialogue evidence. On LoCoMo-10, ProgMem has the highest average F1 with Qwen3-4B, Qwen3-8B, and Qwen3-14B. Its temporal F1 exceeds Mem0 by up to 9.95 points and Zep by at least 12.88 points across these settings. Ablations show gains from route-aware retrieval and two-stage verification. On DialSim, ProgMem improves Answerable and FUQ F1 over Mem0 while using fewer query-time memory tokens. A single-GPU comparison finds higher mean and P95 query latency than Mem0. These results show where state maintenance helps and what it costs at query time.

Multi-agent language systems are often deployed under the assumption that collaboration improves reasoning quality, but this assumption breaks down once token cost, latency, and collaboration-induced errors enter the picture. We study protocol selection under a cost-aware utility objective, where the system decides on each input whether to run a collaborative protocol or fall back to solo reasoning. We propose Causal Advantage Routing for Language agents (CARL), a contextual-bandit framework that estimates the conditional collaboration advantage of each protocol over solo reasoning. CARL combines confounded execution logs with budgeted randomized probes, forms one-vs-solo doubly robust scores and shrunken DR-derived pseudo-outcomes, and at deployment uses a one-sided predictive lower bound for a future realized protocol-solo contrast, otherwise abstaining to solo execution. Across math, code, and multi-hop question-answering benchmarks with an open-weight model, CARL improves cost-aware utility and substantially reduces harmful collaboration over fixed protocols, compute-matched solo strategies, MAS routers, cascade methods, and causal routing baselines, indicating that effective multi-agent systems should estimate when collaboration is causally worth its cost rather than invoking it by default.