AI agents gave companies a cortex, but nobody built the hippocampus. Am I wrong that this is the actual blocker?

Something has been bugging me and I want to check it against people who work with AI every day. A human brain doesn't just know things. It has a part (the hippocampus, roughly) whose whole job is consolidation: taking scattered daily experience and turning it into procedural memory, the "how we actually do this" knowledge. You don't re-derive how to handle an angry customer every morning. Consolidation already turned a hundred experiences into a skill. We now have genuinely capable reasoning (agents as the cortex). We have raw experience piling up everywhere: Slack threads, docs, tickets, that one thread where someone finally explained how refunds actually get approved. Sensory input, tons of it. And we have retrieval/search tools that can find any of it, but that's recall of raw memory, not consolidation. What's missing is the organ that turns the exhaust into consolidated, trusted procedural memory. So every agent deployment I see does one of two things: it improvises ("hallucinated process" is worse than hallucinated facts, because it runs), or someone hand-writes the process docs for the agent, which is just the wiki problem again, and it's stale in six weeks. The knowledge that matters most is exactly the stuff that never reaches a wiki: the workarounds, the exceptions, the "oh we never do X for enterprise customers" tribal rules. It lives in conversation exhaust and people's heads. Humans consolidate it automatically. Companies don't, and agents can't act reliably without it. I'm seriously considering building this missing part. Something that mines "how we actually work" from the exhaust and consolidates it into verified, human-approved procedural memory that any agent can use. But before I sink a year in: Those of you deploying agents at work: is this actually your blocker, or is it something else (integration, permissions, trust, cost)? If you've solved it, how? Hand-curated docs? Karpathy-style markdown wiki? Something like Glean? And would you trust mined process knowledge, or does it only count if a human signed off on it? Genuinely asking. Happy to be told the bottleneck is elsewhere.