@vraj0703/ai-knowledge

Capability registry for AI agents — neural graph, Hebbian strengthening, LSD discovery, manifest-driven adapter runner. Mockable defaults so it runs without SQLite, Ollama, or Langflow.

The "what tools do I have and which one fits this task" layer. Five interfaces (graph store, manifest store, credential store, adapter runner, LSD provider), five default mocks, one DI container that picks per-integration. The manifest store ships with a small generic seed set so search returns sensible results out of the box.

Status: v0.1.0. Lifted from raj-sadan. Mock-default DI runs clone + npm install + npm start in seconds with no backing service.

What it does

| Surface | Purpose | |---|---| | Manifest store | Per-capability TOML manifest (id, name, description, tier, categories) | | Graph store | Nodes (capabilities) + weighted edges (consumer / category / co-usage / lsd) | | Hebbian strengthening | Edges between capabilities used together get reinforced; unused edges decay | | LSD (Latent Semantic Discovery) | LLM finds new edges between nodes that weren't explicitly linked | | Adapter runner | Per-capability TOML manifest names a Node adapter that knows how to invoke the underlying service (Cloudflare API, GitHub, Exa, etc.) | | Credential store | Per-capability secrets (API tokens, etc.) — stays out of the graph |

The pattern: an agent searches for "what tool does X", picks one, invokes it, the act of invoking strengthens edges to other tools that often go together. Over time the graph self-organizes by usage.

Install

npm install @vraj0703/ai-knowledge

Zero dependencies for the default (mock) path. Real path needs three optional peer deps:

• better-sqlite3 — for the real SQLite graph store
• @modelcontextprotocol/sdk + zod — for ai-knowledge mcp

Use

Standalone HTTP service

ai-knowledge serve                       # mocks, port 3489
ai-knowledge --version
ai-knowledge --help

# CLI shortcuts
ai-knowledge search "fetch a URL"
ai-knowledge lookup http-fetch
ai-knowledge register my-tool "My Tool" "Does X for me"

MCP server

ai-knowledge mcp

Four tools auto-discovered:

• knowledge_search — find capabilities by query
• knowledge_lookup — fetch one by id
• knowledge_register — add a new manifest
• knowledge_use — record a capability use (Hebbian reinforce)

From code

const { createContainer } = require("@vraj0703/ai-knowledge/container");

const c = createContainer();              // all mocks
const found = await c.manifestStore.search("llm");
// → matches "ollama-chat" from the seed set

const real = createContainer({ useReal: "graph,manifests" });

Switch to real integrations

KNOWLEDGE_USE_REAL=all ai-knowledge serve
KNOWLEDGE_USE_REAL=graph,manifests ai-knowledge serve

Five known keys: graph, manifests, credentials, adapters, lsd.

Mockability contract

Every external integration ships with a stub:

• InMemoryGraphStore — Map-backed nodes + edges; Hebbian strengthening works in process
• InMemoryManifestStore — seeded with 5 generic capabilities (file-read, http-fetch, ollama-chat, shell-exec, json-extract) so search returns hits out of the box
• StubCredentialStore — returns [mock]-tagged fake secrets, never reads real ones
• StubAdapterRunner — records would-be invocations; nothing executes
• StubLSDProvider — deterministic suggestions based on shared categories (no LLM call)

11 smoke tests run in <150 ms. None touch disk, network, or any LLM.

Configuration

| Var | Default | Purpose | |---|---|---| | KNOWLEDGE_PORT | 3489 | HTTP port | | KNOWLEDGE_USE_REAL | (empty) | Comma-separated integrations to switch to real, or all | | KNOWLEDGE_DB_PATH | data/data_sources/local/knowledge.db | SQLite file | | KNOWLEDGE_OLLAMA_HOST | http://localhost:11434 | LLM provider for direct LSD | | KNOWLEDGE_OLLAMA_MODEL | gemma4:e4b | Model name | | KNOWLEDGE_LANGFLOW_HOST | http://100.108.180.118:7860 | Langflow server (raj-sadan default) | | KNOWLEDGE_HEBBIAN_REINFORCE | 0.1 | Edge weight bump per use | | KNOWLEDGE_DECAY_HALF_LIFE | 30 | Days for unused edges to halve |

What's not here yet

• Generic seed manifests are 5 placeholders. raj-sadan has 71 specific manifests (Cloudflare zones, Exa API key, GitHub OAuth, etc.) that stay raj-sadan-side via KNOWLEDGE_MANIFESTS_DIR.
• Vector similarity search — current search is text matching. Adding an embedder + vector store (could share ai-memory's) is a v0.2 follow-up.
• Live integration tests — only mocks tested. Real SQLite + Ollama + Langflow are exercised by raj-sadan's boot smoke via KNOWLEDGE_USE_REAL=all.

See also

• ai-mind — the cognitive layer ai-knowledge serves
• ai-memory — long-term state, paired with ai-knowledge for context
• ai-senses
• ai-constitution

License

MIT.