Storage and memory

intellibrix defines three small storage interfaces in core. Each has an in-memory implementation for tests and prototypes, and durable ones in separate packages.

Interface	Used for	In memory	SQLite	SQL / Postgres
ConversationMemory	Chat history per thread	MemoryConversationMemory	SQLiteConversationMemory
KeyValueStore	Brick state, caches	MemoryKeyValueStore		SQLKeyValueStore
VectorStore	Semantic search	MemoryVectorStore	SQLiteVectorStore	PgVectorStore

Conversation memory

Pass memory to Intelligence and a threadId to ask. Each thread is its own conversation.

import { SQLiteConversationMemory } from '@intellibrix/memory-sqlite'

const intelligence = new Intelligence({
  provider,
  model,
  memory: new SQLiteConversationMemory({ path: './chat.db' }),
})

await intelligence.ask('I prefer tabs', { threadId: 'user-42' })

@intellibrix/memory-sqlite uses Node’s built-in node:sqlite, so there is nothing native to build. Pass path: ':memory:' (the default) for a throwaway database, or database to share one connection between the memory and vector stores. tableName is validated as a plain SQL identifier. The interface is three methods, append, getMessages and clear.

Key-value stores

import { SQLKeyValueStore } from '@intellibrix/sql'
import { Sequelize } from 'sequelize'

const kv = new SQLKeyValueStore({ sequelize: new Sequelize(process.env.DATABASE_URL!) })
await kv.set('settings:42', { theme: 'dark' })
await kv.get('settings:42') // { theme: 'dark' }

Values are stored as JSON text with a parametrized upsert, in intellibrix_kv unless you pass tableName. It works with any dialect Sequelize supports.

Vector search

Embed text with a provider that implements embed, then store and query vectors:

import { MemoryVectorStore } from 'intellibrix'

const store = new MemoryVectorStore()
const [vector] = await provider.embed!(['Refunds take 5 business days'])
await store.upsert([{ id: 'faq-1', vector, metadata: { text: 'Refunds take 5 business days' } }])

const [query] = await provider.embed!(['how long do refunds take?'])
const matches = await store.query(query, { topK: 3 }) // [{ id, vector, metadata, score }]

score is cosine similarity in every implementation, so you can swap stores without retuning a threshold.

For a persistent store pick one:

// SQLite, no extra services
import { SQLiteVectorStore } from '@intellibrix/memory-sqlite'
const sqlite = new SQLiteVectorStore({ path: './vectors.db' })

// Postgres with pgvector (HNSW index by default)
import { PgVectorStore } from '@intellibrix/sql'
const pg = new PgVectorStore({ sequelize, dimensions: 1536 })

PgVectorStore creates the extension and table on first use. dimensions is required and must match your embedding model, because pgvector columns have a fixed width.

Writing your own

Implement the interface and run the matching contract suite from intellibrix/testing (conversationMemoryContract, and its siblings for key-value and vector stores). If it passes, it behaves like the built-ins.