Plasticity and decay: why good skills rise and stale ones fade

When an agent reaches into the brain, two forces decide what surfaces first: how strong a skill has become, and how fresh it still is. Both are measured rather than asserted, and understanding them is the difference between trusting recall and wondering why a particular skill showed up. This is the most technical guide in the set, and the one that explains why the rest of the system behaves the way it does.

Strength: synaptic plasticity

Skills earn their rank the way connections in a brain do: the ones that get used and pay off get stronger. The critical detail is what counts as "used." It is not how many times a skill was shown, or clicked, or sat in a list. It is real cross-agent reuse, read from the outcome ledger, where a skill was loaded, applied, and reported as having helped.

strength = potentiate(helped, distinctAgents, recency)

The distinct-agent term is doing important work. A skill that three different teammates independently relied on is far stronger evidence of general value than one skill you alone clicked ten times in one afternoon. The former is robust; the latter could be a personal quirk. Ranking by distinct-agent reuse is what makes the brain's top results feel like team consensus rather than one loud voice.

Freshness: mode-selective decay

Strength alone is not enough, because a strong-but-stale skill is dangerous: confident and wrong. So every skill also carries freshness, and freshness is modeled, not a fixed expiry date. Cognition treats recall as a forgetting curve:

R(t) = exp(-(t / S)^β)        // S = 1 / λ₁(L)

The phrase "mode-selective" is the insight. Not everything inside a single skill ages at the same rate, so it does not all decay together:

• High-frequency detail decays fastest: the exact command flags, version numbers, file paths, the specific API shape. These are the parts most likely to be wrong six months out.
• Low-frequency schema stays near-invariant: the underlying principle, the why, the shape of the problem. A migration gotcha's specific command changes; the reason it is a gotcha rarely does.
• When a skill ages past confidence, it does not silently assert stale facts. It surfaces with a prompt to refresh, so you confirm or update before relying on it.

How the two interact

Strength and freshness are not the same axis, and the interesting behavior is at the corners. A strong, fresh skill is the gold standard and surfaces first. A weak, fresh skill is a promising newcomer that needs reuse to prove itself. A strong, stale skill is the dangerous one, valuable history but possibly wrong now, so it asks to be refreshed rather than trusted blindly. A weak, stale skill is a retirement candidate, and decay will quietly let it go.

What you can actually do with this

You do not have to compute any of this; it runs underneath recall. But you can inspect and steer it:

• "show me our skill tree" returns the taxonomy with freshness bands, the strongest skills, suppress candidates, and the next best action.
• "does this skill work?" smoke-tests a specific skill's structure, conciseness, and freshness before you rely on it.
• Reporting outcomes honestly (helped or not) is the lever that moves strength. Refreshing a flagged skill resets its freshness clock.

The payoff for all of this machinery is one property that pure stores cannot offer: the brain gets sharper as it grows, because bad guidance is worse than none. Decay retires what stopped being true while plasticity promotes what keeps working, so usage improves recall instead of burying it.