Build Self-Improving Agents with Memory, Traces, and Skills

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Foundations of continual learning for agents (35 minutes)

• Presentation: What continual learning means for agentic systems; why frozen model weights make it an architecture problem rather than a training problem; key terms (token space, representation space, in-context learning, traces, skills, self-improving agents); where continual learning sits inside the agent harness; why token-space learning is real learning—the “learning without training” result showing in-context learning acts as an implicit low-rank weight update; a brief look at the weight-space frontier (SEAL, sparse memory fine-tuning, nested learning); where the token/weight boundary is dissolving; the taxonomy of token-space versus weight-space improvement and how to decide which lever fits which problem
• Q&A
• Break

High-signal traces and the workflows-to-skills pipeline (45 minutes)

• Presentation: Capturing workflow trajectories with agent memory; canonical trajectory signatures; separating high-signal traces from noise on both the success and failure side; the dual-track promotion pipeline—eligibility gates (frequency, success rate, recency, query diversity); LLM distillation of successful procedures into skill files and repeated failures into guardrail entries; turning heavily used tools and MCP servers into skills; trust-aware skill injection; the threat model for self-improvement (memory poisoning, skill hijacking, adversarial traces laundered into skills); shadow review before activation; drift monitoring and demotion
• Hands-on exercise: Capture traces from live agent runs, inspect trajectory classes including a failure class, run a promotion cycle, review the distilled skill and its paired guardrail, and activate both into the agent’s context

Continual learning across the memory system (40 minutes)

• Presentation: Feedback and update loops for persona, entity, and associative memory, plus handling staleness and conflict as memories evolve; measuring whether learned behavior improves outcomes using five axes (adaptivity, retention, generalization, efficiency, safety); retention framed as skill interference; context failure modes to monitor (brevity bias, context collapse)
• Hands-on exercise: Implement update loops for persona and entity memory, resolve a staleness conflict between an old and a new memory, and wire the full self-improvement loop end to end
• Q&A