Track changing preferences

import { FishMem } from "@fishmem/sdk";

const fishmem = new FishMem({
  apiKey: process.env.FISHMEM_API_KEY!,
});

Let one extraction call turn each change into canonical records with temporal and state-slot fields:

await fishmem.memories.add({
  content: "Sam lives in Lisbon.",
  infer: true,
  user_id: "sam",
  metadata: { source: "profile-conversation" },
});

const moved = await fishmem.memories.add({
  content: "Sam moved from Lisbon to Berlin in May 2026.",
  infer: true,
  user_id: "sam",
  metadata: { source: "profile-conversation" },
});

When structured derivation is enabled, FishMem projects the extracted records into a current-value view with source_ids. It does not keep a second raw canonical copy. Retain the original conversation in your application if it is required for audit or compliance.

Use memories.update(id, ...) only to correct a specific stored record, not to hide a real-world change. Inspect the record's audit trail with:

const history = await fishmem.memories.history(moved.results[0].id);

This distinction keeps “we learned the old record was wrong” separate from “the world changed later.”