memory.
every week the engine reads its receipts and extracts one or two rules. rules are durable. thresholds drift. this is the compounded wisdom of the system — what it knows about itself that it didn’t know before.
shape
every memory rule has three parts. they are always written in this order. the order matters because it forces the engine to prove the rule before it writes the rule.
01
observation.
what actually happened, across many trades. a tendency, not an anecdote.
02
pattern.
what structure the observation has. why the thing keeps happening.
03
rule.
what the engine should do about it. testable, bounded, reversible.
mem-0122026 · wk 16regimeactive
observation
during persistent chop the engine’s best setups were the ones it almost rejected — entries with a slightly lower conviction score than its recent average were the ones that later ran.
pattern
in low-instability regimes, high-conviction signals are disproportionately late. the tape has already moved by the time the confidence pattern resolves.
rule
do not raise the conviction bar in quiet regimes. raise it when instability is rising and everyone is eager to press.
mem-0112026 · wk 15exitsactive
observation
closed trades held longer than the median did not earn more — they paid back the gains during the final leg before the engine’s exit logic fired.
pattern
patience is not a modifier. it is a cost. the engine treats the decision to hold as a second trade, with its own expected value.
rule
every bar of additional hold is a re-entry at mark. if the setup that got you in would not get you in now, exit now.
mem-0102026 · wk 14correlationactive
observation
the book opened three positions on the same beta one morning. they moved as one. when the worst of them hit its stop, the other two were already underwater.
pattern
correlated exposure is not diversification. three trades on one factor is one trade with three names.
rule
correlation penalty applies at portfolio allocation, not at signal approval. a strong signal on coin B is worth less if coin A is already in the book.
mem-0092026 · wk 13rejectionsactive
observation
the most instructive data in a trading week is not the trades. it is the ones the engine refused. the rejections that later ran without the engine told it what it was missing.
pattern
rejections are signal, not noise. a rejection that would have won is a gate tuned too strict. a rejection that would have lost is a gate doing its job.
rule
every rejection is written with the same detail as a trade. the weekly thesis reads rejections first.
mem-0082026 · wk 12calibrationactive
observation
when feedback weights moved too fast, gates became unstable and the engine oscillated. when they moved too slow, gates fossilized against a dead regime.
pattern
any feedback system can feed on itself. a short window amplifies noise. a long window ignores shifts.
rule
calibration must be safety-bounded. no parameter may move more than its authorized delta per window, ever. validator catches it before the risk gate does.
mem-0072026 · wk 11costsactive
observation
the spread on low-liquidity perps was eating a visible fraction of the edge on fast trades. the best signal with a wide book loses to the second-best signal with a tight one.
pattern
cost-adjusted expected value is the only kind of expected value. everything else is a fantasy billed in basis points.
rule
the cost of entry is a first-class feature, not an afterthought. books wider than a threshold get their own sizing penalty before any other modifier fires.
mem-0052026 · wk 09oracleretired
observation
the llm pass-rate never rejected anything. every proposal that passed through it was approved, regardless of quality. it amplified losers as often as winners.
pattern
a gate that never fires is not a gate. it is a ritual.
rule
any gate with a pass rate above a threshold is disabled pending evidence that its rejections correlate with outcome. the oracle layer was disabled in 2026-04.
chain
memory is the input to evolution
rules written here do not live here. they are fuel. once a week the strategist reads the active rules and asks: is there a diff in the engine that would make one of these rules unnecessary? when the answer is yes, the proposal shows up on /evolution. when it merges, the rule is retired. the engine no longer needs to remember it because the engine now is it.