Writing
Posts here, and pieces published elsewhere — shorter and less formal than the papers.
Posts
- Memory as Middleware for Self-Improving AI Agents Agent memory today is trapped inside the agent that produced it. Our Middleware 2026 paper argues it should be a layer instead.
Elsewhere
Written for IBM Research’s blog on Hugging Face.
- ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model Learned memory carrying a non-frontier model to the top of a customer-service agent benchmark.
- Your Agent Aced the Task. Will It Do It Again? On measuring whether an agent is reproducible, not just accurate on average.
- How Much Memory Does Your Agent Actually Need? Agentic memory across eight models — a dose to calibrate, not a feature to switch on.
- Thinking of ACE? We Can Do It with Fewer Tokens Comparing ALTK-Evolve against ACE on AppWorld across the same base models.
- ALTK-Evolve: On-the-Job Learning for AI Agents Introducing ALTK-Evolve — agents that learn from their own past trajectories, with no weight updates and no human annotation.