A split from neuroscience (cortex vs hippocampus) is the best explanation I've found for why AI agents fail on real company work

There's a split from neuroscience I can't stop thinking about as the real reason AI agents fail inside companies. Treat it as an analogy, not a literal claim, but it keeps holding. Your brain runs two memory systems (Complementary Learning Systems theory, McClelland et al. 1995). The neocortex learns slowly and holds general, world knowledge. The hippocampus learns fast: it captures specific episodes as they happen, then consolidates the ones that recur into durable, reusable procedure. A pretrained LLM basically is the neocortex. It read the internet and holds the world's general knowledge. What it does not have is a hippocampus: the fast, company-specific memory that watched how your team actually handled a refund last spring and turned that into a repeatable procedure. So you drop this brilliant cortex into your company and it improvises, and improvised automation fails in production. The real procedure was never in the help doc anyway. It lives in your team's conversations, a couple of people's heads, and one exception everyone now quietly copies. This also explains why the usual tools don't fix it. Retrieval and search are only half a hippocampus: they recall a document but don't consolidate scattered episodes into the real procedure, and the document is often confidently wrong. Agent platforms make you run their agent on their stack. The version of a fix I keep landing on: connect read-only to the tools a team already uses, mine how work actually happens (including the exceptions nobody wrote down), and consolidate the recurring episodes into cited, human-approved, versioned "skills" existing agents could run over MCP, with a human sign-off on anything sensitive. Governance (citations, approvals, an audit trail) has to be the point, because "your AI issued a refund, under whose authority?" is the question that stops people cold. Where I want the pushback: * Is "the agent doesn't know our actual procedures" the real blocker for you, or is it something else (trust, security, the work just isn't repetitive enough)? * Would you connect read-only access to your team's conversations and documents to get this, or is that a hard no? * If you have shipped agents on real workflows, what made them trustworthy enough to turn on? Genuinely hoping some of you tell me where this falls apart.