Give a coding agent project memory

Use agent_id for the agent and run_id for the repository. Together they keep conventions from one project out of another.

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

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

const scope = {
  agent_id: "codex",
  run_id: "repo:billing-service",
};

Recall before work

const { results } = await fishmem.memories.search({
  query: task,
  ...scope,
  top_k: 8,
});

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

Inject projectMemory as context, not as user-authored instructions. Treat memories as potentially stale evidence and allow current repository files to override them.

Remember after work

Store only conclusions that will matter in another session:

await fishmem.memories.add({
  content:
    "Billing API handlers live in src/handlers and export a default async function.",
  infer: false,
  ...scope,
  metadata: {
    kind: "convention",
    source: "architecture-review",
  },
});

Good coding memories include architectural decisions, stable commands, repository conventions, and hard-won gotchas. Avoid transient task state, generated logs, or information that already has a clear canonical file.

FishMem Desktop automates this loop for Codex and Claude Code through a Skill and the local CLI.