Grounded retrieval gets the right documents in front of the model. Context management is what makes sure the right documents are being asked for, by the right person, at the right point in the conversation.
What it actually means
Context management is the layer that maintains three kinds of state:
- Conversation state. What the user has asked so far, what the model has answered, what got clarified, what got refused. A 12-turn conversation should not feel like 12 separate cold-starts.
- Role scoping. Who is the user — a customer, an internal agent, a partner, an auditor? Each role has different retrieval permissions, different tone, different escalation paths.
- Tenant memory. What does the AI remember about this specific customer / account / workspace, across sessions? And — critically — what must it never remember across tenants?
What we build
- A conversation-state store keyed per session, with retention rules driven by your data policy (not the model vendor's defaults).
- Role-based prompt + retrieval scopes — the same question from a customer and an internal ops user routes to different documents and surfaces different answers.
- Federated tenant memory that lets a single mesh app remember things for Tenant A and Tenant B without ever leaking between them.
- Memory-write gates so the model can't surreptitiously persist sensitive content into a long-lived memory.
The failure mode this prevents
The classic context-management failure is the bot that "remembers" something from the last customer's conversation and casually drops it into the next one. Or the support bot that gives an internal-only diagnostic answer to a customer because it didn't know who it was talking to. Or the AI that builds up a year of conversation history and starts contradicting itself because nothing is scoping retrieval anymore.
Context management is the part of Foundations that keeps the AI behaving like a competent employee — knowing who it's in a room with, what they're allowed to hear, and what was just said.