Streaming Continual Learning
AAAI Bridge 2027

A bridge is a cross-disciplinary AAAI event that brings together diverse research communities to explore shared goals and spark new collaborations, with a focus on education, outreach, and community building. The first edition united Continual Learning, Streaming Machine Learning, and Time Series Analysis. This second edition adds Agentic AI, addressing how deployed AI systems can reliably continue learning as their actions reshape the world around them.

Motivation

Learning from data streams is challenging: distributions shift over time (concept drift), models must adapt continuously without forgetting past knowledge, and temporal dependencies are often ignored. Three communities address pieces of this puzzle in isolation: Continual Learning (CL), Streaming Machine Learning (SML), and Time Series Analysis (TSA). But none provides a complete answer. And all three share a common blind spot: they assume the learner passively observes a stream it does not influence.

Deployed AI agents break this assumption. An agent acts on its environment, and those actions reshape the distribution it must then learn from. Non-stationarity becomes partly self-caused. Agents also address knowledge retention very differently from CL: via context windows, retrieval, and episodic memory(often as text) rather than via weight updates. Neither community has a principled account of when each approach is appropriate.

1st AAAI Streaming Continual Learning Bridge aimed to unify CL, SML, and TSA. This second edition extends that vision to agentic AI, the natural, long-overdue fourth thread.

The Bridge

This is the 2nd Streaming Continual Learning Bridge at AAAI-27, Montréal, February 16–17, 2027, building on the first edition at AAAI-26 (Singapore, ~60 attendees, 9 accepted papers). The defining addition for this edition is agentic AI: long-running deployed agents on evolving environments are exactly the realistic, high-stakes setting in which CL, SML, and agent-memory methods each fall short on their own.

The bridge welcomes researchers at any level working where these communities intersect. Register via the AAAI website. The program includes hands-on tutorials (CapyMOA for SML/CL; adaptation and memory in agent systems), a poster spotlight session, paired invited talks, and interactive roundtable and panel discussions. A student mentorship session runs in parallel throughout the day.

We will encourage participants to reason about key open questions, like:

• Can models adapt quickly to new information (SML) without forgetting previous knowledge (CL)?

• What does avoiding forgetting mean under real drift, where decision boundaries shift in previously observed regions?

• Is the loss of plasticity seen in CL also present in SML, and can insights from each field help the other?

• Can temporal dependencies in data streams be leveraged to improve learning?

• Endogenous drift. When an agent's own actions shift the data distribution, what is being detected, and what should adaptation optimise?

• Parametric vs. non-parametric retention. Agents retain past knowledge through context and retrieval; CL retains through weight updates. When should each be used, and can they be composed?

• Drift detection as self-monitoring. Can an agent detect that its environment has shifted and trigger its own adaptation?

• Evaluation under agency. Can prequential evaluation extend to multi-step, sparse, delayed feedback? Which bounded-resource SML results transfer to large pretrained models?