AI Memory and Curated Context
What is institutional knowledge in the context of AI agents?
A senior account manager knows, without checking any documentation, that a particular client will escalate immediately if contacted outside a specific window — the result of an incident three years ago that nearly ended the relationship. She also knows which product configuration that client will never accept, why, and who on their side to call if something goes sideways. None of this is in the CRM. None of it is in the account playbook. It is in her.
When she leaves, so does all of it. Unless someone thought to ask.
Institutional knowledge is the accumulated, unwritten understanding of how an organisation actually operates — the reasoning behind the rules, the context behind the exceptions, the history that explains why things are the way they are rather than some other way. It is distinct from documented policy, which records decisions. Institutional knowledge records the texture around decisions: the conditions under which they apply, the cases they were designed for, the cases they weren't.
Why Agents Expose the Gap
Institutional knowledge has always been a problem for organisations. People leave, take their context with them, and new people spend months reconstructing what was lost. But the impact was bounded — by the time someone joined and got up to speed, the gap was at least partially filled through observation, conversation, and the slow accumulation of experience.
Agents don't work that way. An agent deployed to handle client accounts doesn't spend six months absorbing context. It starts from what it was given. If what it was given is the CRM and the account playbook, that is the totality of what it knows. It applies the documented rules correctly and misses every exception, every nuance, every piece of context that the account manager held and never wrote down.
The agent doesn't know it's missing anything. It's operating competently within the knowledge it has. The failures are quiet — decisions that are technically correct and operationally wrong, in ways that only become visible when the client escalates, or the relationship frays, or someone who knew better reviews the output and recognises the pattern.
What Makes it Hard to Capture
The challenge with institutional knowledge isn't retrieval — it's surfacing. The knowledge exists, in people's heads and accumulated experience. The difficulty is that the people who hold it rarely know what specifically needs to be captured, because to them it is simply how things work. It becomes explicit only when someone asks the right question.
This is why standard documentation processes fail to capture it. A policy review asks "is this policy still accurate?" Institutional knowledge review asks "what do you know about this policy that isn't written in it?" Those are different questions, and most organisations only ask the first one.
The Departure Problem
The gap becomes acute at moments of transition. When the account manager moves on, the question "what did she know that we need to preserve?" is usually asked too late, too broadly, and without a structured way to capture the answers.
The exit interview covers role responsibilities. It asks what projects are in progress, who needs handover notes, what's unfinished. It is designed to transfer tasks, not reasoning. The three-year-old client incident that explains why a specific contact window exists isn't a task to hand over — it's a piece of interpretive history that only surfaces if someone knows to ask for it. Most exit processes don't know what they don't know, for exactly the same reason agents don't.
Agents compound the problem because they distribute the work that previously concentrated institutional knowledge in specific people. When a single person handled an account, the knowledge gap at departure was visible and bounded. When an agent handles the account and the knowledge that informed its behaviour was never captured, the gap is invisible — the agent continues operating, correctly by its documented brief, without any signal that something important is missing.
Engramic's approach
How Engramic Approaches it
There are multiple ways to surface and preserve institutional knowledge. This is one approach.
Engramic's agent onboarding process uses structured questioning to surface knowledge that documentation processes miss. Not "what are the rules for this account?" but "what would you want an agent to know that isn't in the playbook?" — and then working through the answers systematically, narrowing from domain to scenario to specific decision point, until the reasoning behind the exception is captured in a form the agent can draw on. The knowledge doesn't transfer automatically. But with the right questions, it transfers before the person who holds it is gone.