The agent that remembers
how you work.
UKB is the platform's personal-context layer: it records what each user does, detects repeating patterns in that behavior, and turns them into reusable playbooks and a per-user expertise model — so agents act as if they remember. Because they do.
Read the docsAgents without memory make users re-explain the job every session.
Building "agent memory" yourself means per-user behavioral storage, a pattern-mining pipeline over tool-call logs, a suggestion lifecycle with dismissal suppression, embedding infrastructure for similarity matching, and strict tenant isolation across all of it. UKB packages that as one service: consumer apps ask "what does this user know how to do?" instead of building the machinery.
Behavior log
Every tool call a user makes — its inputs, its outcome — is recorded against their account, building a per-user history the platform can mine later.
Pattern mining
A pipeline scans each user's tool-call history for sequences that repeat, independent of exact wording or which session they happened in.
Similarity matching
Embedding infrastructure compares actions and candidate patterns by what they do, not by exact string match, so near-duplicate flows still count as the same pattern.
Suggestion lifecycle
A pattern earns a suggestion only once it crosses a repeat threshold — then it's offered back to the user as a reusable playbook, not auto-applied.
Dismissal suppression
Reject a suggestion once and UKB stops offering it — dismissed patterns are suppressed, not re-served every session.
Tenant isolation
Every user's behavioral data, patterns, and playbooks are scoped to their own account; nothing is mined or matched across tenants.
Expertise model
Accepted playbooks accumulate into a per-user model of what they know how to do — the thing agents actually consult.
Playbooks
A trusted pattern becomes a reusable playbook: a portable description of a task an agent can run on the user's behalf next time.
From one repeated action to a trusted playbook.
THE PATTERN-DETECTION LOOP
Record — the user's action — the tool called, its inputs, its outcome — is logged against their account.
Scan — the system looks for repetition across the user's past actions, not just the current session.
Match — repeated sequences are compared against patterns already known for this user.
Threshold — a pattern must repeat enough times before it's trusted enough to act on.
Suggest — once trusted, the pattern is offered back to the user as a reusable playbook.
Accept — the user, or an agent acting on their behalf, confirms the suggestion — reinforcing it for next time.
Dismiss — a rejected suggestion is suppressed rather than offered again and again.
Knowledge entries and playbooks — the account-scoped API surface.
Built to learn without ever mixing up whose behavior it's learning.
Tenant isolation — per-user behavioral data never crosses accounts; every read and every mined pattern is scoped by the caller's own JWT.
Threshold-gated — a pattern isn't suggested until it's repeated enough to be trustworthy, so playbooks aren't built on noise.
Dismissal-suppressed — a rejected suggestion is suppressed, not re-offered — the system remembers what a user said no to.
Portable playbooks — playbooks are stored as plain descriptions of a task, not tied to a specific tool version or model.
Every decision attributed — each accept or dismiss is recorded against the user who made it, so the learning trail stays inspectable.
What it looks like when an agent actually remembers.
A user pulls the same query and posts it to the same channel every Monday morning. After a few repeats, UKB offers it back as a one-tap playbook.
A user dismisses a suggested playbook a single time. UKB stops offering it — instead of re-surfacing the same suggestion every session.
A user closes their laptop mid-task. When they come back the next day, the agent already knows their usual next step — because the pattern was learned, not re-explained.
A new team member's early actions look nothing like their teammates' established patterns yet. UKB builds their profile from action one — isolated from everyone else's data.
A user calls the same action "close the loop" one week and "wrap it up" the next. UKB matches them as the same pattern by what they do, not by exact wording.
Stop asking users to re-explain how they work.
UKB records, mines, and packages behavior into playbooks and a per-user expertise model — so agents act like they remember, because they do.