Build a support agent with memory

Scope every call by customer:

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

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

Write a resolved conclusion

After a ticket is resolved, let your support workflow produce a self-contained durable summary:

await fishmem.memories.add({
  content:
    "Customer uses the Enterprise plan in AWS eu-central-1; rate limit was raised under ticket T-3391.",
  infer: false,
  user_id: "cust_8842",
  metadata: {
    ticket_id: "T-3391",
    category: "rate-limits",
  },
});

infer:false stores that exact text with zero LLM calls. Keep the source ticket ID in metadata so an operator can audit or correct it.

Recall before answering

const { results } = await fishmem.memories.search({
  query: question,
  user_id: "cust_8842",
  top_k: 5,
});

const knownContext = results
  .map((item) => `- ${item.memory}`)
  .join("\n");

Tell the model not to ask for information already present in knownContext, but also not to treat recalled memory as more authoritative than current account data.

Privacy boundary

The same user_id must be used on write and recall. Enforce the customer ID on the server from authenticated account context; never trust a browser-supplied scope without authorization checks.