Classification Engine

Why Classification Matters

A role memory system that stores everything without structure is not much more useful than a shared drive. The Classification Engine gives every memory entry a type, a domain, and a confidence level — turning raw captured content into structured, searchable, and retrievable institutional knowledge.

Classification is what makes it possible to ask operationally specific questions: "What decisions were made about vendor contracts in this role?" or "What workarounds exist for the month-end reconciliation process?" Without classification, these questions require manual search. With it, they can be answered by the RAG Engine in seconds.

The Memory Taxonomy

RolegacyAI uses a structured taxonomy for classifying role memories. The primary types are:

• Decision: A choice made by the role holder that has ongoing consequences — a vendor selected, an architecture approved, a policy adopted, a trade-off accepted.
• Lesson: Something the role holder learned through experience — what worked, what failed, what would be done differently. Lessons are the institutional equivalent of earned wisdom.
• Workaround: A documented bypass for a known broken or imperfect system, process, or dependency. Workarounds are disproportionately valuable to successors because they are almost never written down.
• Process pattern: A repeatable operational approach that the role has developed — how a particular task is done, in what order, with whom, and why.
• Key relationship: A relationship that has operational significance — a vendor contact, an executive sponsor, an internal ally, a critical dependency.
• Technical configuration: System settings, integration configurations, or tooling setups that are non-obvious and would take a successor significant time to rediscover.
• Risk and escalation context: Known risks, escalation paths, and the conditions under which normal process should be bypassed or escalated.

Multi-Label Classification

A single memory entry can carry multiple classification labels. A decision about a vendor contract may also document a key relationship and a lesson about negotiation timing. The Classification Engine assigns all applicable labels, with a confidence score for each, rather than forcing a single type.

Confidence and Human Override

Every classification is assigned a confidence score. Entries below a configurable threshold are flagged for human review through the Human Validation Loop rather than being committed to the memory store automatically. Role holders and subject-matter experts can correct, relabel, and annotate classifications — improving the model's accuracy over time and ensuring the institutional memory remains trustworthy.

Domain Assignment

In addition to type classification, each memory entry is assigned to one or more knowledge domains relevant to the role — for example: finance and procurement, technology architecture, delivery management, vendor relations, team leadership, or regulatory compliance. Domain assignment is the input to Coverage Scoring, which tracks how well each domain is documented across the role's accumulated memory.