Client Brain

Client Brain stores tenant-scoped memories extracted from eligible resolved tickets. Those memories can add typed context to the initial ticket-agent run. Learning and recall are both guarded, so a stored memory does not show that it changed a response or improved an outcome.

Current Scope

Client Brain currently supports four memory types:

Type	Stored shape
Business	A non-empty business fact
Voice	Supported tone and style fields
Issues	An issue pattern, best resolution, and frequency
Process	A trigger and action

Typed validation prevents unsupported memory shapes from entering the agent prompt. Historical or malformed rows can still appear in stored-row statistics, but they are not admitted as typed prompt context.

Learning Pipeline

1. Eligible terminal ticket

Learning can be requested when a ticket newly enters a supported terminal state: solved, closed, or resolved. It is not a continuous interaction-observation stream.

2. Runtime checks

Before extraction, the learner re-reads the exact tenant ticket and checks that it is still eligible. It can skip when the ticket is missing, no longer terminal, or the required runtime provider key is unavailable.

3. Memory extraction

Eligible tickets can produce tenant-scoped business, voice, issue, or process memories. The current producers use the ticket, its latest classification, the latest outbound non-note message, and relevant approval evidence. Knowledge-base content, CSAT ratings, canned responses, and generic response ratings are not direct Client Brain training sources.

4. Initial-run recall

The initial ticket-agent run fetches candidates in two bounded tiers inside the finite 0.5–1 confidence 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. If there is no overlap, the confidence-ordered fetched set is the deterministic fallback. Copilot's memory tool also excludes confidence outside the finite 0–1 domain before its 30-row limit. After typed admission, the initial prompt can include up to 30 business facts, one formality voice memory, five issues, and five process rules; valid business facts over the cap are counted, and the prompt reports how many were omitted. The continuation path can reload fresh tenant-scoped live Client Brain context when eligible rows are present, and retrieval is not a semantic relevance search. Configured-provider execution and behavioral improvement are not yet available as a customer-facing result. Very common words (such as "the", "for", "you") are ignored when counting overlap.

Viewing Stored Memory

Go to Insights → Analytics → Client Brain.

The current dashboard:

• loads an initial page of up to 100 tenant-wide stored rows;
• provides Load more memories for subsequent bounded offset pages;
• restarts from the first page when business, voice, issues, or process is selected;
• reports tenant-wide typed rows, all stored rows, average valid stored confidence, invalid-confidence rows, recorded activity, and stored counts by type even when the row list is filtered;
• labels unsupported values as legacy memory format.

The Typed Rows aggregate exactly mirrors the shared four-family value guard without fetching tenant values into the application. It is not a count of memories a particular prompt will use: confidence, the mutable 30-candidate window, formality selection, and the 8/1/5/5 prompt caps apply separately. Stored and typed-row totals retain invalid historical confidence for visibility, while the confidence average includes only finite 0–1 values and a separate warning reports excluded confidence rows. Recorded activity is a mixed contribution/selection counter and does not show that a memory affected an answer.

Pages are ordered by recorded activity and then confidence. Offset paging is not a stable snapshot, so concurrent learning or activity changes can shift rows between requests; the dashboard removes duplicate row IDs when appending.

The dashboard does not currently provide editing, deletion, or reset controls. Client Brain mutation endpoints are suspended.

Retention and Isolation

Client Brain reads are tenant-scoped. Cross-tenant contribution and global bootstrap are suspended while deterministic privacy, provenance, withdrawal, and evidence controls are completed.

Experiments

Experiments are a separate feature. Client Brain does not automatically select experiment winners or turn experiment outcomes into memories.

Learn about the current Experiments boundary →

FAQ

Does Client Brain learn from all resolved tickets?

No. An eligible terminal ticket can start a learning attempt, but extraction can skip at runtime or produce no memory.

How quickly will it learn my style?

There is no guaranteed learning timeline. Stored voice memory depends on eligible evidence and successful extraction, and its effect on deployed behavior still needs to be checked in the product.

Can I edit, delete, or reset memories?

Not currently. The dashboard is read-only and mutation endpoints are suspended.

Does the brain forget?

Not automatically at present. Pruning is suspended, so retention must not be described as active decay or a guaranteed time window.

Does browser automation train Client Brain?

No automatic browser-playbook or authentication-flow learning is part of the current Client Brain contract.