@aletheia-labs/adapters-openai

OpenAI Responses-compatible adapter for Aletheia authority-governed memory.

Status: 0.1.0 public baseline. Reference OpenAI integration is live as a narrow library adapter. It does not own OAuth, API keys, terminal UX, tools, publication, or provider account state.

Requirements

• Node 20+.
• ESM-only. Add "type": "module" to your package.json, use .mjs, or use a build tool/runtime that handles ESM. CommonJS require() is not shipped in 0.1.x.

Quickstart

pnpm add @aletheia-labs/core @aletheia-labs/store-sqlite @aletheia-labs/adapters-openai openai

import OpenAI from 'openai';
import { AletheiaAuthority, staticVisibilityPolicy } from '@aletheia-labs/core';
import { AletheiaOpenAIResponsesBridge } from '@aletheia-labs/adapters-openai';
import { openSqliteStores } from '@aletheia-labs/store-sqlite';

const stores = openSqliteStores('./aletheia.sqlite');
const authority = new AletheiaAuthority({
  eventLedger: stores.eventLedger,
  memoryStore: stores.memoryStore,
  conflictRegistry: stores.conflictRegistry,
  visibilityPolicy: staticVisibilityPolicy([{ kind: 'private:user' }]),
});

const bridge = new AletheiaOpenAIResponsesBridge({
  client: new OpenAI({ apiKey: process.env.OPENAI_API_KEY }),
  authority,
  eventLedger: stores.eventLedger,
  model: 'gpt-5',
});

Hosts own credentials, retries, rate limits, and provider selection. This adapter only receives an already-authenticated client.

What this package does

• Records conversation turns as append-only Aletheia events.
• Asks an OpenAI Responses-compatible client to draft memory proposals as JSON.
• Sends every draft through AletheiaAuthority.propose() before anything becomes a MemoryAtom.
• Recalls memory through AletheiaAuthority.recall() and re-checks action authority through tryAct() before calling the model for an answer.
• Refuses to call the model when recall/action authority fails closed.

Client contract

The package does not import the OpenAI SDK. It accepts any caller-provided object with responses.create(input):

const bridge = new AletheiaOpenAIResponsesBridge({
  client: new OpenAI({ apiKey: process.env.OPENAI_API_KEY }),
  authority,
  eventLedger: stores.eventLedger,
  model: 'gpt-5',
});

What this package does NOT do

• No OAuth, device login, refresh-token handling, or ChatGPT subscription plumbing.
• No OpenAI SDK dependency.
• No authority upgrade from model output.
• No tool execution.
• No semantic retrieval, embeddings, vector index, or ranking.
• No bypass around propose, recall, or tryAct.

Fixture Demo

After building the workspace:

pnpm -r run build
pnpm -F @aletheia-labs/adapters-openai run demo:fixture

The fixture uses a fake Responses client and should print boundaryViolations: [].

Stability

Public surface for the initial library cycle:

• AletheiaOpenAIResponsesBridge
• OpenAIResponsesClient
• ConversationIngestionInput
• AnswerWithRecallInput
• ConversationIngestionResult
• AnswerWithRecallResult

Everything else is adapter plumbing and may change during the 0.x line.

Development

From the repo root:

pnpm install
pnpm -F @aletheia-labs/adapters-openai typecheck
pnpm -F @aletheia-labs/adapters-openai test
pnpm -F @aletheia-labs/adapters-openai build
pnpm -F @aletheia-labs/adapters-openai run demo:fixture