Agent memory that learns from wins is grading the user, not the context

The interesting failure is not the missed document. A miss produces a bad outcome and gets fixed. The one that survives is the wrong material sitting beside a task that succeeded anyway, which then looks useful in the record afterward ^[1]. What the system does with that record is credit the document for work the person did ^[4].

Where the coin flip sits in the code decides whether it means anything. In ZCL the decision happens in the context provider, which creates the session id, decides whether the request should explore, asks value attribution for learned document values, handles cold start, and filters selected document ids through causal evidence before the bundle goes back to the agent; assignment happens ahead of that handoff ^[13]. A flip applied after selection would only relabel a choice the user's judgement had already made.

The price shows up in the sample budget. The default minimum sample threshold is 20, and it rides on the hypothesis object itself so the brake travels with the proposed change ^[10]. The generator builds order hypotheses from the top five documents already used for a task type and inclusion tests for up to three uncertain documents ^[11]. Three inclusion tests at twenty observations each is sixty sessions of a single task type before those questions can resolve, and one hundred and twenty if twenty is the count per branch rather than the total ^[15]. A deployment with real traffic absorbs that. A deployment serving one developer's agents spends months on three questions.

The author is direct about the ceiling. ZCL is described as his own context learning platform, and the post is offered as an account of how much the layer can currently prove, which he says is less than its own vocabulary suggests ^[8]. Whether the minimum sample threshold is connected to anything that blocks promotion is deferred to a later section ^[14]. That deferral is where the load sits. A metric declared in advance disciplines the analysis honestly. A threshold that no code enforces is a comment with a number in it.

None of this is a published result, and one repository is not evidence that randomized provisioning produces better agents. The diagnosis travels further than the implementation: when a memory system's only teacher is the record of tasks that went well, its teacher is the users who were going to do well ^[4].