How TeamAI Turns Experience into Searchable Context

TeamAI’s context layer answers a narrower question than “Can an agent search the web?” It asks whether an agent can recover the right team experience—and know when that experience is incomplete—inside the scope of the workspace where it is working. This chapter follows that context from a new learning or codebase fact through indexing, recall, voting, and maintenance. The practical value is not a promise that the top result is correct; it is a way to make team memory available while keeping relevance and uncertainty visible.

A useful context system has to solve two opposing problems. If it stores only a few curated notes, it may miss the detail needed for a difficult incident. If it exposes every document to every project, the signal-to-noise ratio falls and private project experience can leak across boundaries. TeamAI addresses the tension by separating where learning lands from how the local index is built, then combining broad retrieval with explicit scope and evidence signals.

stateDiagram-v2 [*] --> Created [*] --> CodeExtracted state "Learning or code fact" as Created state "Structured code extraction" as CodeExtracted state "Scoped source" as Scoped state "Local index built" as Indexed state "Ranked recall" as Recalled state "File or source read" as Read state "No hit or weak coverage" as Gap state "Vote or revision feedback" as Feedback state "Reported or maintained" as Maintained Created --> Scoped CodeExtracted --> Scoped Scoped --> Indexed Indexed --> Recalled Recalled --> Read : result worth checking Recalled --> Gap : no result or missing terms Read --> Feedback Feedback --> Maintained Maintained --> Indexed : revise or rebuild Gap --> Created : document missing context

The important loop is visible in the transitions. Creation is not the same as visibility; a learning or code fact must enter the scope’s index. Recall is not the same as acceptance; a result should be read and checked. Feedback can improve ranking or produce a better document, but a retrieval failure remains a signal that knowledge is missing, not an invitation to invent an answer.

From a contribution to a ranked memory

In the ordinary Git-backed path, a contribution is a Markdown learning placed in the team repository’s learnings/ area. The document can carry structured metadata such as a title, author, date, and tags. When a member has one active logical project with a learning namespace, the contribution is routed there; with no applicable namespace, it goes to the shared root. That rule gives the document an ownership context before it ever enters search.

The next transition depends on scope. In user scope, TeamAI mirrors the team’s learning files into the user’s local TeamAI learning directory. In project scope, it can index the learning files in the team repository directly. Pull rebuilds a local index after synchronizing the relevant resources, and a successful contribution also rebuilds the index immediately so the new document is recallable without waiting for another session start. If a network push fails, the contribution is saved outside the disposable team clone for a later retry; a temporary failure does not turn a locally created lesson into an unrecoverable change.

The index is deliberately broader than the learnings/ directory.