Asked where the recursive self-improvement loop starts to fall over — whether errors compound, whether exception handling constrains it — Tan locates the failure precisely: “once you get a lot of markdown files, or yeah, anytime you have like some large amount of data, that’s where like provenance matters. That’s where conflict management — when you have two facts, if you have provenance then you’re like, well, this is the more recent one, this is the one that wins. You need a cron job that’s sort of sweeping through that and making sure that things are accurate and from the right places.”
Three things follow, and they are worth separating. Recency is the tie-breaker, which means every fact needs a timestamp and a source, not just a value. Accuracy decays on its own, so the sweep has to be continuous rather than a one-time validation. And the constraint appears only at scale — a handful of skill files never surfaces it, which is exactly why it arrives as a surprise.
This is the failure mode that Evolving summaries beat append-only memory — rewrite profiles, don't accumulate facts is designed around, seen from the other side: evolving a summary is one way of resolving conflicts, but it presumes you can already tell which version supersedes which, and provenance is what supplies that. It bounds Domain-specific skill libraries are the real agent moat, not core infrastructure with a maintenance cost — an accumulated library is a moat only while its contents remain trustworthy, and an unswept library quietly becomes a liability. The same caveat applies to Persistent agent memory preserves institutional knowledge that walks out the door with employees, where preserved knowledge that nobody reconciles preserves the contradictions too.
In his Startup School keynote Tan states the same constraint as a warning rather than a prediction: “a brain nobody curates is a garbage dump with great search.” Retrieval will surface a stale fact with total confidence, and a bad skill file encodes a bad process forever, so the primitive he names is memory plus hygiene — provenance on every fact, contradiction checks when new information collides with old, and “a librarian whose actual job is pruning.” Treat the brain like production infrastructure and it compounds; treat it as a dumping ground and you get “a very confident agent that is wrong in ways nobody can trace.” That is the curation half of Retrieval is the primitive, not the product — being worth retrieving from is the product, and the reason the capture discipline in Compound engineering makes each unit of work improve all future work needs pruning attached to it rather than capture alone.