intent-brain

qmd-team-intent-kb

A governed team memory platform for Claude Code powered by qmd.

Turn ephemeral Claude Code session insights into persistent, governed, team-wide memory. qmd-team-intent-kb captures institutional knowledge generated during AI-assisted development, applies deterministic governance policies, and makes curated knowledge searchable through qmd's local full-text indexing.

Part of the Compile-Then-Govern stack — INTKB is the govern layer. Upstream, intentional-cognition-os (compile) emits the spool it consumes; downstream, qmd (retrieve) serves the curated result with qmd:// citations. → Ecosystem overview

Positioning

This project has a clear separation of responsibilities:

• qmd is the local search and index engine. It does one thing well: fast, offline full-text retrieval.
• qmd-team-intent-kb provides everything around it: orchestration, governance, lifecycle management, deduplication, analytics, and the team-shared memory model.
• Enterprise-capable from day one. Tenant isolation, audit trails, and governance policies are architectural requirements, not afterthoughts.

Architecture Thesis

The system is built on a strict information flow with deliberate trust boundaries:

1. Claude Code proposes memory. During sessions, Claude Code generates memory candidates — insights, patterns, decisions, and lessons. These are proposals, not facts.

2. Deterministic code decides what becomes durable team memory. A policy engine evaluates every candidate against configurable governance rules. Secret filtering, deduplication scoring, relevance evaluation, and tenant isolation are all enforced programmatically. No LLM judgment is involved in promotion decisions.

3. qmd serves as the local retrieval and index edge engine. Curated, approved memories are synced to local qmd indexes for fast, offline search. The edge daemon handles replication and conflict resolution.

4. A canonical memory control plane owns truth, lifecycle, governance, deduplication, and analytics. The control plane API is the single source of truth for memory state. It manages the full lifecycle: inbox, review, promotion, archival, and supersession.

5. Git is a mirror, export, and distribution layer — not the canonical database. The git exporter publishes curated knowledge to git repositories in structured Markdown + frontmatter format. Git enables distribution and versioning, but it is downstream of the control plane.

6. Default search targets curated knowledge only. When a developer or Claude Code session queries memory, the default search scope is curated, policy-approved knowledge. This is a deliberate design constraint.

7. Inbox and archive content must not pollute default search. Raw candidates in the inbox and retired memories in the archive are excluded from default retrieval. Accessing them requires explicit, intentional queries. This prevents unvetted or stale content from contaminating team knowledge.

Monorepo Structure

qmd-team-intent-kb/
├── apps/                    # Deployable applications
│   ├── api/                 # Control plane REST API (Fastify)
│   ├── curator/             # Memory promotion, dedupe, supersession
│   ├── edge-daemon/         # Local qmd sync daemon
│   ├── git-exporter/        # Git mirror/export
│   └── reporting/           # Analytics and lifecycle dashboards
├── packages/                # Shared libraries
│   ├── schema/              # Shared Zod schemas and domain types
│   ├── qmd-adapter/         # qmd CLI/API integration layer
│   ├── claude-runtime/      # Claude Code session capture
│   ├── policy-engine/       # Memory governance policy evaluation
│   ├── repo-resolver/       # Multi-repo context resolution
│   └── common/              # Shared utilities
├── kb-export/               # Git-exported curated knowledge output
├── 000-docs/                # Project documentation and RFCs
├── examples/                # Usage examples and templates
├── scripts/                 # Build and maintenance scripts
├── tests/                   # Integration tests (repo root)
└── .github/                 # CI/CD workflows and templates

Applications