@anvia/memory-drizzle: Usage Patterns

Exported table definitions

Use drizzleMemorySchema when you want all memory tables:

export const schema = {
  ...drizzleMemorySchema,
};

Use individual table exports when your schema file groups tables manually:

export { agentMemorySessions, agentMemoryMessages, agentMemoryErrors };

Tenant-scoped conversations

const memory = createDrizzleMemoryStore(db, {
  scope: { metadataKeys: ["tenantId"] },
});

The default scope includes sessionId and userId; metadata keys add product tenancy or workspace isolation. The adapter reads selected keys from context.metadata, so the matching session call must pass the same values:

const session = agent.session(conversation.id, {
  userId: user.id,
  metadata: { tenantId: user.tenantId },
});

Set includeUserId: false only for intentionally shared tenant memory. Scope chooses which database rows to read and write; your app still owns authorization.

Load initial messages

Use session.messages() on the server when a React page needs to render an existing thread:

export async function GET(request: Request, params: { threadId: string }) {
  const user = await requireUser(request);
  const agent = createSupportAgent(user);

  const messages = await agent
    .session(params.threadId, {
      userId: user.id,
      metadata: { tenantId: user.tenantId },
    })
    .messages();

  return Response.json({ messages });
}

The response contains core Anvia Message[]. Convert those messages before passing them to useChat:

import type { Message } from "@anvia/core";
import { initialMessagesFromMemory, useChat } from "@anvia/react";

export function SupportChat({
  threadId,
  messages,
}: {
  threadId: string;
  messages: Message[];
}) {
  const chat = useChat({
    endpoint: `/api/threads/${threadId}/chat`,
    initialMessages: initialMessagesFromMemory(messages),
  });

  return <ChatProvider controller={chat}>{/* thread */}</ChatProvider>;
}

initialMessages hydrates browser state only. The stream route should still call agent.session(...).prompt(latestUserMessage) so Drizzle memory loads authoritative history on the server.

Advisory locking

Append and failed-run writes use Postgres advisory locks by default through db.execute(...). Set lock: "none" only when your deployment serializes same-scope writes another way.