Associative Recall

Decision sharing across projects. Every project’s decisions are private by default. When a platform team’s decisions need to be visible to every product team, or two related projects want their corpora connected, a project owner grants an explicit share. The grant is audit-tracked, revocable with a captured reason, and never bypasses your organization’s zero-trust posture. It lives on a Decision Sharing tab on the project detail page: grant, revoke, or view received grants in one place.

Cross-project relationships. An organization-wide setting decides whether the relationship engine considers only the source decision’s own project (the safest option, and the default) or widens to include projects that have granted a share into it. Same seven relationship types, same UI, same governance: SUPERSEDES, REQUIRES, CONSTRAINS, CONTRADICTS, EXTENDS, IMPLEMENTS, RELATED_TO. Cross-project rows carry a colored chip so the source is unmistakable. Find it under Settings, Configuration, Relationship Scope.

The organization decision graph. A new top-level Decision Graph opens a canvas of every decision your team can read across every project in the organization. Cross-project synapses are drawn separately from within-project ones, so the places where your architecture crosses team boundaries are visible at a glance. The view respects the same zero-trust visibility rules as the rest of the platform, and does not depend on the Relationship Scope toggle.

An Architecture Context API, over a standard protocol. External AI coding tools can now read your decision corpus through two standard interfaces. A REST API at /api/context/ exposes four operations: search, single-decision fetch, scope constraints, and a primary relevance bundle. An MCP server at /api/mcp exposes the same four as MCP tools: search_decisions, get_decision_with_relationships, get_constraints_for_scope, and relevant_context. Configure Claude Code, Cursor, GitHub Copilot, or ChatGPT and Codex once, and the agent working in your codebase can ground its suggestions in what your team actually decided. Every response respects your organization’s zero-trust visibility. Every call is authenticated by a per-organization API token that only your admins can issue, revoke, and rate-limit.

Outcome-weighted ranking. When an agent asks for the decisions most relevant to what it is working on, the ranking now reflects what your team learned, not only what looks similar. Decisions whose implementations were validated get a boost. Decisions the team explicitly abandoned or flagged for revision are penalized. Decisions sitting on unresolved critical feedback are downweighted. Similarity remains the primary signal; the outcome adjustment surfaces alongside it in the response, so the agent can reason about why a particular decision is being surfaced. This is what separates a recall layer from semantic search over your decision records. Search returns what looks similar. Recall returns what your organization learned.

The MCP integration this release ships is the same one our own development sessions now use to ground code changes in our captured architectural decisions. The Recall API is the interface every AI agent working on Gateway Architect calls before proposing a change. The corpus your team builds is the working memory of the tools building alongside you.

That is why every future release compounds. The more decisions your team captures, the sharper the ranking, the richer the constraints your tools see, and the tighter the loop between what you decide and what gets built. Associative Recall (Stage 3) is where that loop leaves the product and enters the tools you already work in.