Jake Broekhuizen distinguishes stored traces from memory an agent can use on a future run. Semantic memory covers facts and preferences, episodic memory covers examples and experiences, and procedural memory covers instructions and skills. Working context is temporary, while long-term context guides later behavior.
Jake Broekhuizen describes a read-write cycle: load relevant context, capture an agent's trajectory, filter the evidence for useful signal, then update durable context. A financial-services assistant that slips from measured guidance into directive language illustrates why observation alone does not correct future behavior.
Jake Broekhuizen maps the loop onto LangSmith observability, background analysis and a context hub. Most trace data should remain history or evaluation data rather than become memory. He warns that stale runtime caches can hide updates and recommends human review before changing high-impact instructions or policies.
Watch on YouTube
