Keva Brain
Stateless AI processes support requests in isolation, forgetting what happened before. Keva Brain provides a guarded tenant-scoped memory path.
Tenant-scoped typed memory can add eligible support context to an initial agent run. Configured, deployed behavioral improvement remains to be proven.
Traditional AI support agents can treat a repeat request like a blank slate. A customer writes in about the same billing issue for the third time, and the earlier context is unavailable.
Durable records are useful only when the right evidence is admitted safely. Client Brain provides one typed, tenant-scoped context path for eligible support memory.
Keva can make selected tenant-scoped memory and evidence available as agent context.
Typed recall is available to the initial agent run
Recall is guarded and optional. Continuation can reload fresh tenant-scoped live Client Brain context.
Ticket arrives
A customer submits a support request through email, widget, or API.
Initial recall
The initial agent run can load valid typed tenant memory. Continuation can reload fresh tenant-scoped live Client Brain context.
Agent receives context
Admitted memory is prompt context, not proof that it changed the response or outcome.
Eligible learning attempt
A newly terminal ticket can request extraction; runtime and provider gates can skip it.
Coral steps show the Client Brain boundary. Learning is requested only for eligible terminal tickets and can skip; prompt admission is limited to valid typed memory.
Four typed memory categories
Business, voice, issue, and process values must pass typed validation before prompt admission.
Layer 1: Business Memory
“Refund requests require a valid receipt.”
A non-empty business fact can be extracted from eligible terminal-ticket evidence and stored with source-ticket provenance.
Business fact: Refund requests require a valid receipt. Source: eligible terminal ticket
A valid business fact can be included in a later initial agent run. Inclusion is not proof that it changed a response.
Layer 2: Voice Memory
“Use a formal, concise support style.”
Supported voice fields can be extracted when eligible evidence includes a human-edited AI draft and qualifying outbound response.
Voice: Level: formal Prefer phrases: "Thank you" Avoid phrases: "No problem"
Voice memory can provide style context to an initial run. Its effect on the generated response requires product verification.
Layer 3: Issue Memory
“Shipping delays recur after a particular fulfilment state.”
An issue memory requires a pattern, a best-resolution description, and a finite frequency value.
Issue: fulfilment delay after dispatch Best resolution: verify carrier scan, then update customer Frequency: 3
Issue memory can expose a prior pattern and resolution as context; it does not guarantee that the model will reuse it.
Layer 4: Process Memory
“Verify identity before issuing a refund.”
A process memory stores a trigger and an action extracted from eligible approval and resolution evidence.
Trigger: customer requests a refund Action: verify identity before issuing the refund
Process memory can remind an initial run of a stored trigger/action pair. Following it remains a model behavior to verify.
Browser actions are not Brain training
Browser automation is a separate execution capability. Client Brain does not automatically learn reusable browser playbooks, cache authentication flows, or turn navigation sessions into typed support memory.
Explicit Playbooks
Defined and supervised separatelyA browser workflow can be defined and supervised as a separate playbook. Successful execution does not automatically create a playbook or prove that later runs will be faster.
Execution Evidence
Kept separate from Client BrainBrowser actions can produce execution records and outcomes. Those records must not be described as automatic Client Brain learning, reusable authentication memory, or verified navigation improvement.
Current boundary: Browser automation and Client Brain are separate. Reuse, speed, and reliability improvements require configured, deployed behavioral verification.
Context available, behavior still to verify
These are context boundaries, not guarantees that a model used the memory or improved an outcome.
No durable tenant memory is available
A valid typed memory can be included in an initial run
Prior support evidence is unavailable
Stored evidence can be presented as additional context
Recent failure context is absent
Available safety evidence can accompany the request
Repeat-customer context is absent
Structured customer context can be made available
Platform context must be rediscovered
Previously recorded connector context can be supplied
Guarded, typed, and tenant-scoped
For the technically curious. Keva Brain is designed to be invisible.
Skip-capable learning
Learning can skip when ticket eligibility, runtime, or provider gates are not met. A skipped attempt must not be counted as a learned outcome.
Typed initial context
The initial run fetches candidates in two bounded tiers inside the finite 0.5–1 range. The fallback tier fetches up to 200 candidates ordered by confidence, newest row, then ID. When the ticket subject or latest customer message has words of at least three characters, a second tier also fetches up to 200 candidates whose memory key or value contains any of those words case-insensitively, so a matching memory is not lost to the fallback window when a business has thousands of stored rows. It keeps voice memories first. For each non-voice memory, it counts shared normalized Unicode words or numbers of at least three characters between the ticket subject or latest customer message and the memory key or value; positive overlap ranks first, then higher confidence, then the original fetched order. With no overlap, the confidence-ordered fetched set is the deterministic fallback. It then admits up to 30 business / 1 renderable voice / 5 issue / 5 process rows; valid business facts over the cap are counted, and the prompt reports how many were omitted. Continuation can reload fresh tenant-scoped live context.
Source provenance
Learning re-reads the exact tenant ticket and stores source-ticket references. Multiple memories from one attempt are replay-safe individually, not one all-or-nothing transaction.
Read-only dashboard
The authenticated dashboard lists bounded offset pages and separates shared-guard-valid typed rows from all stored rows. Prompt selection remains narrower. Editing, deletion, and reset controls are not currently available.
Retention custody pending
The registered pruning job is suspended until tenant retention and deletion custody is implemented. No automatic decay, expiry, or deletion guarantee applies.
Tenant-scoped
Client Brain reads and learning custody use tenant identity. Cross-tenant contribution and bootstrap remain suspended pending stronger privacy and provenance controls.
Frequently asked
When does Keva Brain run?
Learning is attempted only for eligible solved, closed, or resolved tickets and can skip when runtime or provider gates are not met. Valid typed memory can be loaded for an initial agent run; continuation can reload fresh tenant-scoped live context.
How much does Brain add to my AI costs?
Brain is included in eligible plans with no separate Brain fee. Actual provider usage depends on successful extraction attempts; no fixed per-ticket cost or latency claim is made.
Can I see what the Brain knows?
Yes. The authenticated Client Brain dashboard separates typed-value-valid rows from all stored rows, supports type filters, and can load subsequent bounded pages. Typed-row totals are not a claim that every row enters a prompt. The dashboard is currently read-only: edit, delete, and reset controls are not available.
What happens if I disconnect a platform?
Connector knowledge is deleted automatically (CASCADE on connector deletion). Action journal entries are retained for historical reference since they're linked to tickets and customers, not the connector itself.
Does Brain share data between tenants/workspaces?
Per-tenant Brain data stays tenant-scoped. Cross-tenant global contribution and bootstrap are suspended while deterministic privacy, provenance, withdrawal, and evidence controls are completed.
Does Brain automatically forget old memories?
No active decay or pruning guarantee applies. The registered pruning job is suspended until tenant retention and deletion custody is implemented.
An AI agent that actually remembers.
Eligible terminal tickets can contribute tenant-scoped memory. Behavioral improvement requires product verification.
Custom setup tailored to your stack. We adapt to you.
Book a Demo