Where your AI learns your business
Plain-English business definitions. Schema-free playbooks. Conventions that survive every tool change.
UKB is the prefrontal cortex of the agent. It holds what your organization means by 'an active customer,' how a renewal check should be run, who must approve a refund, and the unwritten conventions that turn a generic agent into yours. UKB stays in plain English so humans always win the conflict — and so the knowledge survives every infra swap underneath.
What lives in UKB
Knowledge stays schema-free so it survives provider swaps, column renames, and migrations.
Why prompt instructions are not a knowledge base
Three reasons 'put it in the system prompt' stops working past a single use case.
A custom instruction either fires every turn or it doesn't. Real business knowledge applies to specific entities, roles, and moments.
'Search Notion' is a time bomb. When you switch document stores, every system prompt across your org breaks at once.
Auto-memory accumulates facts but has no structure for resolving conflicts. The one rule that mattered gets smoothed away in compaction.
What UKB does
Four properties that make business knowledge first-class.
Plain-English business definitions
Define 'active customer,' 'Q3 ARR,' 'support escalation' in English. The agent reads them at planning time and threads them through every step.
Schema-free playbooks with domain tags
Playbooks read 'search [document_store],' 'post to [team_chat].' Swap providers and the playbook keeps working — the binding happens against your SKB-connected account.
Conflict precedence — humans always win
When admin-written knowledge contradicts discovery or usage-learned facts, UKB applies a deterministic precedence: humans > admin discovery > usage. Nothing gets quietly overridden.
Survives compaction by design
UKB entries are structured records, not unstructured memory. The Orchestrator pulls them at planning time, so they never get lost in a summary.
Three sources, one source of truth
How UKB grows without devolving into chaos.
Admins write the canonical definitions
Your team writes business definitions, playbooks, and conventions in plain English. These outrank every other source.
Discovery proposes new knowledge
When the agent encounters a recurring pattern it can't resolve, it drafts a UKB candidate and queues it for admin review. Nothing trusted without approval.
Usage tightens the rough edges
Repeat user corrections and confirmed patterns flow back as low-priority UKB updates — always below admin-written knowledge.
When sources disagree, humans always win.
Knowledge base, honestly compared
What UKB gives you vs. the alternatives.
| Capability | Interactor UKB | ChatGPT / Cowork | Build from scratch |
|---|---|---|---|
| Knows your org by definition | Yes — 3-layer profiles | Project context only | Yes — 3+ months |
| Playbooks with domain tags | Yes — portable | Prompts hard-coded | Yes — 4+ months |
| Survives infra changes | Yes — schema-free | Breaks on rename | 6+ months careful |
| Learns over time, humans win conflicts | Yes — precedence | No structured learning | 6+ months |
| Time to a working knowledge base | Days | N/A | 6–9 months |
Bring a definition. Leave with a UKB draft.
Pick one business term that you wish your AI understood. We'll structure it as a UKB entry and show you how it changes the plan.
