Is RAG the same as agent memory?

RAG retrieves passages that match a query. Memory also has to track time: which decision is newer. Archilas keeps the date on every source and cites it. See the post on memory versus RAG.

Does a larger context window replace memory?

No. It holds more of one session, but it is empty at the start of the next one, and models can still miss facts that are in the prompt.

Blog / 2026-10-30 · 2 min

What AI agent memory is, in four parts

By [OPERATOR NAME]

Four objects in a row on a dark table, standing for the parts of agent memory

AI agent memory is anything that lets an agent use information from outside the current prompt. In practice there are four kinds: the context window, instruction files, a memory store the agent queries, and shared team memory built from the tools a team already uses. They solve different problems.

1. The context window

Everything in the current conversation. It is exact while it lasts and gone when the session ends. A bigger window does not fix this, and putting evidence in the prompt does not guarantee the model uses it. The preprint has numbers on that for a small reader.

2. Instruction files

CLAUDE.md, AGENTS.md, Cursor rules. Someone writes them and the agent reads them at the start. They are good for stable conventions and bad for anything that changes weekly, because a person has to keep them current.

3. A memory store your code queries

Libraries and services such as Mem0, Zep, and Letta give developers an API to save and retrieve memories inside an application they build. That suits a product that remembers its own users. There is a sourced overview in how Mem0, Zep, and Letta store memory.

4. Team memory

What the team decided, built from the places the work happens: Slack, GitHub, Gmail, Notion, Linear, Calendar, Drive, and AI chats. Archilas gives AI a memory of the team. Every agent a teammate runs can ask it over MCP and get the current decision with its source and date. Slack is one of those sources, not the product.

Which one you need

Repeating the same fixed rules: an instruction file.
Building an app that remembers its end users: a memory store with an SDK.
Your own agents keep missing what the team already decided: team memory.

Questions

Is RAG the same as agent memory?: RAG retrieves passages that match a query. Memory also has to track time: which decision is newer. Archilas keeps the date on every source and cites it. See the post on memory versus RAG.
Does a larger context window replace memory?: No. It holds more of one session, but it is empty at the start of the next one, and models can still miss facts that are in the prompt.

All posts