Company Context Maturity Model.
Updated 2026-06-07 · Agent-readable markdown available
A practical model for moving from scattered work history to governed company context for AI agents.
Teams get leverage from agents when useful work becomes reusable context with source, scope, review, and reuse.
- The jump is from keeping records to changing the next agent run.
- Each level adds a control agents need: source, scope, freshness, conflict review, and audit.
- Ambience gives teams a path from first memories to context operations without forcing a large connector project first.
Level 0: cold starts
Agents begin every session as if the company has no memory. Humans paste context, repeat decisions, and hope the agent notices the important parts.
This is where most teams start. It feels workable while usage is small, then breaks as soon as multiple people and agents touch the same work.
Level 1: personal recall
One person has prompts, notes, local files, or private memory that helps their agent behave better. The context is useful, but it does not travel safely to teammates.
The risk is accidental authority. A private preference starts to look like a team rule because the agent has no better source.
Level 2: project memory
Important decisions and conventions are attached to a project. Agents can retrieve the current implementation constraints before they edit, plan, or respond.
Ambience makes this level tangible with project-scoped memories, source links, redaction state, and session-start context.
Level 3: source-linked team memory
Calls, tickets, PRs, docs, threads, and agent sessions become approved source material. The memory is the durable takeaway, not the whole artifact.
This is the first compounding level. A Granola call can inform a Codex session. A GitHub PR can inform a future Linear plan. The source remains inspectable.
Level 4: governed company context
Context is shared across the organisation with scopes, redaction, conflict review, access changes, and audit. Agents receive context because they are allowed to see it, not only because it matches a query.
This is where Ambience becomes infrastructure: memory reads and writes are visible product events rather than invisible prompt text.
Level 5: context operations
Teams run lightweight context ops: weekly reviews, source sweeps, stale-memory cleanup, skill promotion, and context readiness checks before large agent rollouts.
The outcome is not a perfect archive. It is a living company context map that keeps agents close to the work the business actually trusts.