What Is AI Memory? A Practical Guide to Persistent Memory for Agents

Ask most people what "AI memory" means and they'll point at the scroll-back in their chat window. But conversation history isn't memory — it's a transcript that vanishes when the session ends. A bigger context window isn't memory either; it's a bigger short-term buffer that still resets to zero next time. Persistent memory is something else: a durable, structured store of what an agent has learned about a subject, that survives across sessions, tools, and time.

The three things people confuse with memory

• Context window — the tokens the model can see right now. Large, but ephemeral and re-paid on every call.
• Chat history — a log of past turns. Retrievable, but unstructured and siloed to one tool.
• RAG — retrieval over static documents. Great for facts in a corpus; it doesn't learn about you.

Memory is the layer that turns raw activity into durable knowledge: it decides what's worth keeping, distils it into a structured form you can query, and carries it forward. The test is simple — close the session, open a new one in a different tool, and ask a question that depends on last week. If the answer is there, that's memory.

What a real memory layer does

A persistent memory layer runs a loop: observe, learn, recall. It ingests observations, distils them into a knowledge graph, and makes them retrievable with provenance — so a recalled fact comes with where it came from. The good ones also decide what not to keep, so your prompts don't balloon with noise.

// The whole loop, in three calls.
await mm.memory.observe({
  subject: { kind: 'contact', externalId: 'sarah' },
  content: 'Prefers email over phone.',
})

const hits = await mm.memory.search('how to reach sarah', { limit: 5 })
// → recalls the fact, with provenance, in a brand-new session.

Why agents specifically need it

Agents are the sharpest case because they act over long horizons and across tools. A coding agent that forgets your architecture re-learns it every session. A support agent that forgets a customer asks them to repeat themselves. Memory is what lets an agent accumulate competence instead of resetting to a blank slate — and, at MemMesh, it's also the substrate for prediction: once you remember a subject's history, you can forecast what they'll do next.

Chat history tells you what was said. Memory tells you what it means — and, if it's any good, what happens next.

If you're evaluating a memory layer, judge it on four things: does it persist across sessions and tools, does it carry provenance, can you govern and erase it, and — the one most tools skip — does it do anything with what it remembers beyond handing it back?