Tacit Fragments: Operationalising Tacit Knowledge as a Governed Memory Layer for Agentic AI

For more than half a century, tacit knowledge has explained why skilled organisational work cannot be reduced to written rules or formal procedures. Competent action relies on situated judgement, perceptual cues, and experience-shaped discrimination. Historically, AI systems could not access this know-how because they required explicit rules, labelled examples, or articulated document corpora. The shift to agentic AI changes this practical reality: agents can now use tools, maintain memory, observe workflow traces, and interact with human workers directly. This capability raises a question: which aspects of tacit practice leave enough behavioural traces and contextual conditions to become useful computational artefacts? This paper presents Tacit as a governed memory layer for agentic AI that complements procedural, semantic, and episodic memory. To operationalise it, we introduce Metis, a reference architecture specifying how systems can responsibly capture, represent, validate, govern, and retrieve "tacit fragments". We delineate an exogenous pathway that captures fragments from situated human practice, the paper's primary focus, alongside an endogenous pathway where agents surface latent competence from their own operational traces. We set out how agentic systems can turn situated human practice into governed memory an agent can use, provided each fragment remains strictly bound to its conditions, provenance, confidence, and human validation.