Architecture Overview

KnowledgePulse is an AI knowledge-sharing protocol built on a dual-layer architecture. The protocol layer defines how knowledge is structured, discovered, and governed, while the infrastructure layer provides the concrete runtime components that implement the protocol.

Protocol Stack

The protocol is organized into five layers, each building on the one below it:

┌─────────────────────────────────────────┐
│            5. Governance                │
│   Licensing, attribution, audit trail   │
├─────────────────────────────────────────┤
│            4. Privacy                   │
│   Scoping, redaction, GDPR compliance   │
├─────────────────────────────────────────┤
│            3. Discovery                 │
│   Search, scoring, vector retrieval     │
├─────────────────────────────────────────┤
│            2. KnowledgeUnit             │
│   JSON-LD schema, types, versioning     │
├─────────────────────────────────────────┤
│            1. SKILL.md                  │
│   Agent capability declaration format   │
└─────────────────────────────────────────┘

Layer 1 -- SKILL.md defines a standard file format for agents to declare their capabilities, inputs, outputs, and constraints. It is the foundational contract that other layers reference.

Layer 2 -- KnowledgeUnit specifies the JSON-LD schema for three knowledge types (traces, patterns, and SOPs) and the versioning rules that govern schema evolution.

Layer 3 -- Discovery covers how knowledge units are indexed, searched, scored for relevance, and retrieved -- including vector-based similarity search.

Layer 4 -- Privacy handles scoping rules, field-level redaction, and GDPR-mandated operations such as data export and the right to be forgotten.

Layer 5 -- Governance addresses licensing, attribution chains, and audit trails that track how knowledge flows between agents.

Component Overview

Four runtime components implement the protocol stack:

Component	Package	Role
SDK	@knowledgepulse/sdk	Core library -- types, capture, retrieve, scoring, SKILL.md parsing, migrations
Registry	registry/	Hono REST API server -- stores knowledge units, handles auth, rate limiting
MCP Server	@knowledgepulse/mcp	Model Context Protocol server -- exposes 6 MCP tools, dual-mode registry bridge
CLI	@knowledgepulse/cli	Command-line interface -- search, install, validate, contribute, auth, security

SDK

The SDK (packages/sdk) is the core library that all other components depend on. It provides TypeScript types generated from Zod schemas, functions for capturing and retrieving knowledge units, relevance scoring, SKILL.md parsing and sanitization, and chainable migration functions for schema version upgrades.

Registry

The Registry (registry/) is a Hono-based REST API server that acts as the central store for knowledge units. It handles authentication via Bearer tokens, tier-based rate limiting, and exposes CRUD endpoints for knowledge units.

MCP Server

The MCP Server (packages/mcp-server) exposes KnowledgePulse functionality through the Model Context Protocol, making it accessible to any MCP-compatible AI client. It operates in dual mode:

• Standalone mode -- runs with its own in-memory store.
• Proxy mode -- bridges to a remote registry via the KP_REGISTRY_URL environment variable.

CLI

The CLI (packages/cli) provides a command-line interface for developers and agents to interact with the protocol -- searching for knowledge, installing SKILL.md files, validating schemas, contributing knowledge units, and managing authentication.

Monorepo Structure

knowledgepulse/
├── packages/
│   ├── sdk/            # @knowledgepulse/sdk — types, capture, retrieve, scoring, migrations
│   ├── mcp-server/     # @knowledgepulse/mcp — 6 MCP tools, dual-mode registry bridge
│   ├── cli/            # @knowledgepulse/cli — search, install, validate, contribute, auth
│   └── sop-studio/     # Placeholder (Phase 3)
├── registry/           # Hono REST API server (in-memory stores, auth, rate limiting)
├── specs/              # codegen.ts, validate-consistency.ts, skill-md-extension.md
├── examples/           # basic-sdk-usage, mcp-client-example, langraph-integration
└── website/            # Docusaurus documentation site

Storage Model

Phase 1 uses in-memory Map<> stores behind async Promise<T> interfaces. This design is intentional: the async interface boundary means the storage backend can be swapped to a persistent database (PostgreSQL, SQLite, etc.) without changing any consumer code. The vector cache uses a brute-force linear scan rather than HNSW, and is likewise swappable via its interface.

Tech Stack

Technology	Purpose
Bun	JavaScript/TypeScript runtime and package manager
Hono	Lightweight HTTP framework for the Registry and MCP Server
Zod v3	Schema definition and validation (with zod-to-json-schema for codegen)
tsup	Build tool for the SDK (ESM + CJS + TypeScript declarations)
Biome	Linter and formatter
@modelcontextprotocol/sdk	MCP protocol implementation
bun test	Test runner (319 tests across 15 files)