nori-memory

TypeScript SDK and trusted Hono HTTP server for Nori memory.

Scope

• Tenant isolation uses orgId.
• Input is asynchronous: inputMemory() only creates memory_inputs and a queued memory_runs record.
• A worker calls processMemoryRun() to extract memory and write resources/items/categories. Category summaries are marked dirty and refreshed by processCategoryRefreshJob().
• Recall returns only shared memory plus the requester member's private memory.
• Production storage is PostgreSQL with pgvector.
• HTTP server does not include authentication. It is intended for trusted backend-to-backend calls.

SDK

import {
  OpenAICompatibleEmbeddingClient,
  OpenAICompatibleLlmClient,
  PostgresMemoryRepository,
  createCloudWatchStdoutTelemetry,
  createNoriMemory
} from "nori-memory";

const memory = createNoriMemory({
  repository: new PostgresMemoryRepository({
    connectionString: postgresConnectionString
  }),
  llmClient: new OpenAICompatibleLlmClient({
    baseUrl: "https://openrouter.ai/api/v1",
    apiKey: extractionApiKey,
    model: "openai/gpt-4.1-mini"
  }),
  embeddingClient: new OpenAICompatibleEmbeddingClient(
    {
      baseUrl: "https://openrouter.ai/api/v1",
      apiKey: embeddingApiKey,
      model: "openai/text-embedding-3-small"
    },
    1536
  ),
  telemetry: createCloudWatchStdoutTelemetry({
    environment: "prod",
    service: "nori_memory"
  })
});

const input = await memory.inputMemory({
  orgId: "org_1",
  actorMemberId: "mom",
  idempotencyKey: "conversation_1:turn_10",
  occurredAt: new Date().toISOString(),
  messages: [{ role: "user", content: "不要在晚上9点以后安排日程" }],
  memberContext: {
    actorMemberId: "mom",
    members: [{ memberId: "mom", nickname: "mary", roleName: "mom" }]
  }
});

await memory.processMemoryRun({ runId: input.runId });
await memory.processCategoryRefreshJob();

If telemetry is omitted, NoriMemory emits CloudWatch Embedded Metric Format JSON to stdout with environment="unknown" and service="nori_memory".

NoriMemory depends on LlmClient and EmbeddingClient only. Owning services can inject OpenAI-compatible adapters, AI SDK adapters, or their own implementations. The standalone server builds AI SDK clients from llm.memory_infer, llm.memory_embedding, and secret config.

createNoriMemory() does not require config when callers inject llmClient and embeddingClient. config now contains runtime strategy only, such as recall and category refresh defaults.

Worker

import { runNoriMemoryWorker } from "nori-memory/worker";

await runNoriMemoryWorker({
  memory,
  pollIntervalMs: 2000,
  reapIntervalMs: 60000,
  categoryRefreshBatchSize: 1,
  batchSize: 1,
  signal
});

HTTP

import { createNoriMemoryServer } from "nori-memory/http";

const app = createNoriMemoryServer({ memory });

Routes:

• POST /v1/memory/inputs
• GET /v1/memory/runs/:run_id
• POST /v1/memory/recall
• POST /v1/memory/items
• PATCH /v1/memory/items/:item_id
• DELETE /v1/memory/items/:item_id
• GET /v1/memory/items
• GET /v1/memory/categories
• GET /healthz

The SDK core never reads environment variables. Load secrets in the owning service and pass repository/client implementations into createNoriMemory().