Agent Card Format

The agent card format is the standard for knowledge units that agents consume via the memory bus. It replaces relational lookup tables with flat files that embed into pgvector through the existing seed pipeline. No new DB schema is required — the growdirect_memory vector store is the only persistence layer.

Design Principles

One concept per card. A card holds exactly one idea — a signal, an agent, a hierarchy axis, a lifecycle gate. If a card needs subsections for multiple concepts, split it.

Every section embeds independently. The memory bus seeder chunks by heading. Each section must make sense when retrieved without the rest of the card. Write sections as if they will be read cold.

Frontmatter is machine data. All relational facts (what feeds what, which agent owns this, which module consumes it) live in frontmatter arrays. The body is prose that embeds well. Agents can filter on frontmatter fields before running semantic search.

No volatile data. Row counts, live metrics, and current-state summaries do not belong in cards. Cards capture structure, relationships, constraints, and context.


Frontmatter Schema

Required fields

Field Type Values
card-type enum signal-feed · agent-profile · org-layer · field-hierarchy · role-binding · domain-module · lifecycle-gate · infra-capability · platform-thesis · runbook · format-spec
card-id string kebab-case unique identifier, stable across versions
card-version integer Increment on any substantive change
domain enum lp · merchandising · finance · labor · platform · local-market · cross-cutting
layer enum domain · infra · local-market · cross-cutting
status enum draft · approved

Optional fields

Field Type Purpose
agent string Agent that owns or primarily uses this card
feeds array card-ids or module codes this concept feeds into
receives array card-ids or module codes that feed into this concept
tags array Search terms for memory bus recall
milestone string M1–M6 PMO milestone (domain modules only)
last-compiled date ISO 8601, set by content-engine lint
needs-review boolean Set by content-engine lint on frontmatter issues

Body Conventions

Section headings (use these in order, omit any that don't apply)

## What this is
One-sentence definition. This is the identity chunk — retrieved first by semantic search.

## Purpose
Agent-facing "why this exists." What problem does it solve, what decision does it inform.

## Structure
For hierarchies, role models, org layers: the data shape, constraints, enumerations.

## Signal
For signal feeds: what the signal contains, frequency, format, quality characteristics.

## Consumers
Which agents or modules use this. What they do with it. Be specific about the action, not just the name.

## Sources
Where the data comes from. External systems, feeds, human inputs.

## Routing
How it flows through the agent network. Show the chain explicitly.

## When to run
**Runbook cards only.** The trigger condition — what state means this procedure is needed. Be specific: "Docker stack is down and dispatch work is about to begin" not "whenever you need to restart Docker."

## Preconditions
**Runbook cards only.** What must be true before starting. Ordered list. If a precondition is not met, the procedure must not proceed.

## Canonical steps
**Runbook cards only.** Exact commands in order, with expected output noted inline. No paraphrasing — copy-paste fidelity is the standard.

## Verification
**Runbook cards only.** How to confirm the procedure succeeded. A query, a health check, a curl — something observable that produces a pass/fail signal.

## Failure modes
**Runbook cards only.** Known failure patterns with recovery steps. Each mode: symptom → cause → fix. No speculation — only confirmed failures.

## Invariants
Hard constraints. What must always be true. What must never happen. Agents enforce these.

## Related
Links to related cards using card-id syntax. One line per related card with a brief note.

How Cards Enter the Memory Bus

Cards in Brain/wiki/ (including subdirectories) are picked up by the seed pipeline:

python3 services/memory-bus/scripts/seed_standalone.py

The seeder chunks each card at heading boundaries and embeds each chunk with qwen3-embedding:8b (1024-dim). Chunks are stored in growdirect_memory with the file path and heading as metadata.

Agent recall:

memory_recall("social threat detection signal local market")
memory_recall("merchant org hierarchy head office role")
memory_recall("geography hierarchy LP district")

Filtered recall (frontmatter fields): The seeder stores frontmatter as chunk metadata. Future memory bus versions will support pre-filter by card-type, domain, or layer before vector search.


Versioning

Increment card-version on any substantive change. The seed pipeline is incremental by mtime — a version bump forces re-embedding even if the file timestamp doesn't change (when using --force). Do not reset card-version to 1 on edits; that signals a rewrite, not an update.


Card Directory

All agent cards live in Brain/wiki/cards/. Subdirectories are permitted for large domains but the flat directory is preferred until there are more than ~50 cards.

Card ID Type
Merchant Org Hierarchy merchant-org-hierarchy org-layer
Geography Hierarchy geography-hierarchy field-hierarchy
Category Hierarchy category-hierarchy field-hierarchy
Role Binding Model role-binding-model role-binding
Local Market Agent local-market-agent agent-profile
Signal: Seasonality signal-seasonality signal-feed
Signal: Weather + Zone SEO signal-weather-seo signal-feed
Signal: Social Threat Detection signal-social-threat signal-feed
Signal: Civil Services signal-civil-services signal-feed
Signal: Community Intelligence signal-community-intel signal-feed
Signal: Property & Landlord signal-property-landlord signal-feed
Infra: Blockchain Evidence Anchor infra-blockchain-evidence-anchor infra-capability
Infra: L402 OTB Settlement infra-l402-otb-settlement infra-capability
Platform Thesis platform-thesis platform-thesis
Platform: Closed-Loop Attribution platform-closed-loop-attribution platform-thesis
Platform: Retailer Lifecycle Test platform-retailer-lifecycle-test infra-capability
Platform: ALX as VSM platform-alx-vsm platform-thesis
Runbook: Memory Bus Seed runbook-memory-bus-seed runbook
Runbook: Docker Startup runbook-docker-startup runbook
Runbook: Brain Wiki Commit runbook-brain-wiki-commit runbook
Runbook: Create a Runbook runbook-create-runbook runbook
Runbook: Vault Publish runbook-vault-publish runbook
Retail: Vendor Lifecycle retail-vendor-lifecycle domain-module
Retail: Vendor Compliance Standards retail-vendor-compliance-standards domain-module
Retail: Vendor Scorecard retail-vendor-scorecard domain-module
Retail: Chargeback Matrix retail-chargeback-matrix domain-module
Retail: Purchase Order Model retail-purchase-order-model domain-module
Retail: Three-Way Match retail-three-way-match domain-module
Retail: Receiving Disposition retail-receiving-disposition domain-module
Retail: Inventory Valuation / MAC retail-inventory-valuation-mac domain-module
Retail: Inventory Audit retail-inventory-audit domain-module
Retail: Backroom Cost Transfer retail-backroom-cost-transfer domain-module
Retail: Merchandise Financial Planning retail-merchandise-financial-planning domain-module
Retail: Demand Forecasting retail-demand-forecasting domain-module
Retail: Replenishment Model retail-replenishment-model domain-module
Retail: AP / Vendor Terms retail-ap-vendor-terms domain-module
Retail: Merchandise Hierarchy retail-merchandise-hierarchy domain-module
Retail: Site Management retail-site-management domain-module
Retail: Event Management retail-event-management domain-module
Retail: Operations KPIs retail-operations-kpis domain-module
Retail: Assortment Management retail-assortment-management domain-module
Retail: Sales Audit retail-sales-audit domain-module
Retail: Import Management retail-import-management domain-module
Retail: Space, Range & Display retail-space-range-management domain-module
Retail: Item Authorization retail-item-authorization domain-module
RaaS: Receipt as a Service raas-receipt-as-a-service infra-capability
Platform: Enterprise Document Services platform-enterprise-document-services platform-thesis
ICP: Murdoch's Reference Implementation icp-murdochs-reference platform-thesis
Platform: Proof Case platform-proof-case platform-thesis
Platform: Wyoming Ecosystem platform-wyoming-ecosystem platform-thesis
Platform: Field Capture platform-field-capture platform-thesis
Platform: Performance NFRs platform-performance-nfrs platform-thesis
Platform: PwC Benchmarks platform-pwc-benchmarks platform-thesis
Store: Network Integrity Monitoring store-network-integrity platform-thesis
Platform: PII Hashing platform-pii-hashing platform-thesis
Platform: Multi-Tier Assortment platform-multi-tier-assortment platform-thesis
Platform: Cryptographic Erasure platform-cryptographic-erasure platform-thesis
Platform: Data Classification platform-data-classification platform-thesis
Platform: Stack Commitment platform-stack-commitment platform-thesis
Platform: Architectural Continuity platform-architectural-continuity platform-thesis
Platform: L402 + ILDWAC Moat platform-l402-ildwac-moat platform-thesis