hardAI Engineering

How do agents maintain memory across steps and sessions?

Q: How do agents maintain memory across steps and sessions?

**TL;DR.** Three layers: **working memory** = the context window (this task, right now); **scratchpad** = files/notes the agent writes so state survives compaction; **long-term memory** = an external store of durable facts across sessions, retrieved when relevant. Hard problems: **selection** (store little), **staleness** (memories age), **injection** (memories are data, not instructions). working context window volatile, per-request <text class='d-text' x='229'

114 views

Understand the problem

Working context, scratchpads, and external long-term stores — what lives where.

memoryagentsstatelong-term-memory

Attempt it yourself

Sketch your approach before reading the solution — that's what interviews test.

Stuck? AI Nudge Available

Get a conceptual hint to guide your logic without spoiling the final implementation.

Study the solution

The solution is waiting

Give it an honest attempt first — then compare your thinking with the full walkthrough.

Read the code

Memory record designed for safe recall

await memory.upsert({
  key: "deploy-process",                       // upsert: update, not append-duplicate
  fact: "Production deploys go through GitHub Actions 'release' workflow; " +
        "manual kubectl is forbidden.",
  observedAt: "2026-07-04",                    // staleness is checkable
  source: "session-8f2c",                      // provenance for audit
});

// recall path treats memories as data:
const mems = await memory.relevant(task, 3);
prompt.push("[Background notes — verify before relying on them]\n" +
            mems.map((m) => "- (" + m.observedAt + ") " + m.fact).join("\n"));

Join the discussion

Discussion (0)

No responses yet. Be the first to share what you think.

What is red-teaming an LLM application and how is it done?