AI Agents of the Week

Research you should know about

In this week’s agent highlights:

1. A clear theme was memory and continual learning. Several papers introduced new ways for agents to accumulate and reuse knowledge across tasks without retraining their base models. One approach (Meta-Policy Reflexion) converts an agent’s textual self-reflections into a persistent memory of rules that improve future decisions. Another (ArcMemo) focuses on storing abstract concepts distilled from past reasoning, enabling test-time learning that compounds over multiple problem-solving sessions. These memory innovations aim to overcome the stateless nature of today’s large language model (LLM) agents, so agents can learn from experience and avoid repeating mistakes.

2. Better planning was another focal point. Researchers proposed a dynamic planning framework that trains agents when to think ahead and when to act directly. By allocating computational “reasoning” steps only when needed, this method makes long-horizon decision-making both more effective and more efficient. In parallel, a reinforcement learning technique showed that honing an agent’s tool-use planning (with fine-grained rewards for each step) can dramatically boost overall task success, underlining the importance of targeted training for the planning module.

3. Agentic edge AI also saw progress. A new retrieval-augmented framework for mobile UI agents demonstrated how combining external knowledge, tool use, and memory can greatly expand an agent’s capabilities on smartphones. It achieved substantial gains on complex multi-step tasks by retrieving relevant app information and recalling successful past strategies.

Taken together, this week AI agents were pushed toward being continuous learners with smarter planning strategies and greater resilience in unpredictable environments. Below we’ll take a look at the research – outlining their core innovations, why they matter for agents in practice, the specific challenges they tackle (memory, planning, robustness, etc.), and the new capabilities or future implications they bring.