roi-zalta walks the context timeline: RAG (2023) → agentic RAG → context engineering (2025) → Karpathy’s Obsidian/LLM-wiki moment (April 2026) → the “company brain”, shared organizational knowledge that agents consume and humans maintain. RAG still covers ~99% of needs; a company brain isn’t solved yet, mostly because keeping sources synced and contradictions resolved is expensive. He proposes four layers (factual memory → human communication / “the missing why” → context-graph ontology → governed actions/loops), plus ontology “lenses” per role, authority weighting, and nightly “dreaming” jobs that consolidate memory. Also: humans as knowledge nodes, agents and humans sharing workspaces and boards, and a hot take that SaaS is not dead.
Show: ai-engineering-podcast
Concepts: second-brain · rag · memory-consolidation · knowledge-graphs · agent-workspaces · open-knowledge-format · wiki-retrieval · ai-guardrails