The memory that travelled too far
Imagine correcting an agent while writing an incident report: “Lead with impact, then show the timeline. Do not open with a celebratory summary.” A week later, the same sentence appears as guidance for a personal project update. The agent remembered the preference, but it did not understand its edge.
The failure is not missing memory. It is missing scope. The remembered text was valid evidence for operational communication, perhaps for one team or one mode. It was not a universal instruction for every piece of writing. Increasing the retrieval window would only make the mistake more likely.
Recall answers the wrong question
Memory systems are good at asking which stored items are semantically similar to a new query. Similarity can find a past incident report. It cannot decide whether the judgment inside that report is allowed to govern the current task. Meaning and authority are related, but they are not identical.
Scoped guidance adds explicit boundaries. A change can belong to a project, a task family, a mode, or a temporary situation. It also has a facet: fact, goal, current state, voice, taste, judgment, workflow, exemplar, anti-pattern, or uncertainty. A horizon says whether it is for one run, a temporary period, ongoing work, or durable use.
These dimensions stop a profile from becoming one undifferentiated biography. A current deadline should expire without erasing a stable boundary. A copywriting anti-pattern should not be treated as an identity declaration. A preferred review workflow can be ongoing while a project goal remains temporary.
Build a pack for the task
Meraki Core uses those boundaries to assemble a MerakiPack: a deterministic set of approved guidance, examples, constraints, and provenance for one run. The pack is deliberately smaller than the full profile. Its usefulness comes from exclusion as much as inclusion.
A manifest records why each item was selected. That makes leakage reproducible. If the incident-report rule appears in a personal update, a person can inspect the task, mode, scope, and matching decision. The fix can be a narrower scope rather than a vague request for the model to be more careful.
This also creates a security boundary. Text retrieved from a document may help answer a question, but it should not gain authority to edit durable profile state. Untrusted content can become evidence for review; it cannot promote itself into active guidance.
The proof still needed
The local Core loop can produce scoped packs deterministically, and its comparison path can test a pack against a baseline, raw memory, and a targeted removal. That is real infrastructure for asking whether scope improved a result.
It is not yet proof that the chosen facets and horizons match how people naturally think, or that a person can diagnose leakage without becoming a profile administrator. Independent evaluation and hosted use remain open. The design could be technically precise and still demand too much attention.
The unresolved question is whether scope can become a calm product experience rather than a settings panel. Can Meraki infer a narrow candidate, explain it in one sentence, and let a person approve or adjust it without learning an ontology? More memory will not solve that. Better boundaries might.