The vocabulary of the context layer

Tribal knowledge is operational knowledge the company uses every day but has never written down where a system can read it: what a metric means here, which of several look-alike tables is official, which exception has applied since when.

a16z used the term in “Your Data Agents Need Context” as the main reason data agents go wrong in production. It is usually conditional (“for CRM, new USCAN deals since 2025 live in Affinity”) and it lives with people, not in a repository: Slack threads, PDFs, a handful of roles. A context layer writes it down as versioned entries, the team approves them, and agents receive them when they answer.

Context drift is the gap between what a governed corpus says and what its sources now contain: the definition that changed in dbt last month, the table that was dropped, the policy that was replaced.

A one-time export goes stale and agents keep answering as if nothing changed. Renbase re-syncs connectors and compares them to their sources. Changed content produces a new draft for review. A source that disappears marks the entry as stale, and the answer reports that. The last-verified date travels with every response so an agent can decide whether to trust the definition or ask again.

Abstention is the system declining to answer. It says “I don't know” and gives the reason, instead of guessing, when context is missing, expired or in conflict.

For a data agent, that is the difference between no number and a wrong number given to a customer or an auditor. It has to come from the architecture, not from a prompt: the system must know what approved sources can support and refuse the rest. In Renbase, abstention is a normal response. The agent is told which context was missing or conflicting, so each abstention is a task for the corpus: add or fix the definition.

A typed citation is a citation with structure: not only a link, but the exact entry or document, its version, who approved it, and when it was last verified against its source.

A plain citation tells you where a claim came from. A typed citation lets you defend it. When someone asks why the agent said that, you can show the entry and the approver. In Renbase every claim in every answer carries one, which is what makes an agent's answer auditable in a regulated environment and reviewable anywhere else.

Provenance is the recorded origin of a piece of context: which connector, document or person produced it, and which versions it has passed through since.

When two approved definitions of “active_user” overlap, returning both with their provenance (finance, reporting scope; product, analytics) lets the team pick. The model does not pick in silence. Provenance is also the audit trail behind every typed citation.

The answer contract is Renbase's rule for every response: always cite, abstain when the corpus cannot support an answer, and return conflicts instead of resolving them quietly.

The name makes the behavior testable: sources always; a person approves before anything shapes an answer; no invented definition. That is what separates a context layer from retrieval infrastructure. Retrieval finds similar text. The contract requires an approved source, or a clear refusal.

MCP (Model Context Protocol) is the open standard that lets AI agents discover and call external tools and context sources. Claude, Cursor and many other agents speak it natively.

For a context layer, MCP is how the corpus is delivered. Any MCP-compatible agent connects to Renbase over Streamable HTTP with a per-organization key and gets five tools (get_definition, search_context, list_entities, expand_source and ask) under the same governance, credits and limits as the organization's people.