tropo

The Agentic Builders / New to Agents · 3 of 14

What is a Context Window (and Why Agents Forget)

· 4 min read

New to Agents · Beginner Path · 3 of 14

An AI agent can feel brilliant for an hour, and then in the next conversation have no idea who you are or what you were building together. That is not a bug or a cheap model. It is the single most important thing to understand about how agents work, and once you see it, almost everything else about Tropo makes sense.

An LLM - the model underneath an agent - does not remember things the way you do. It has a context window: a fixed-size space holding all the text it can see right now. Everything the agent knows in the moment lives in that window - its instructions, your conversation so far, and any files it has read. It reads all of it, top to bottom, every single time it responds.

Two things follow, and they explain almost everything about working with agents.

The window has a fixed size. As your conversation grows and the agent reads more, the window fills up. When it is full, the oldest text falls out of view to make room. The agent is not choosing to forget the start of your conversation - it literally can no longer see it.

The window is wiped when the session ends. The next session starts blank. The model did not learn anything from last time; it simply read what was in front of it then. A fresh session is a brilliant stranger.

So the context window is working memory, not long-term memory. It is the desk the agent works on, not the filing cabinet it remembers from.

The context window: working memory, not long-term memory WHILE THE AGENT WORKS new input CONTEXT WINDOW · FIXED SIZE System instructions Your conversation Files it has read …oldest text falls out when full session ends WHEN THE SESSION ENDS ↻ wiped blank the next session starts here The model did not learn - it read what was in front of it, and the box is emptied every time.

Why this matters for Tropo

If working memory gets wiped every session, then anything you want an agent to keep has to live somewhere outside the window - written down, on disk, where the next session can read it back in.

That is the whole reason Tropo is built on plain markdown files and a navigable vault: durable memory the agent reloads at the start of every session, so the next conversation does not begin as a stranger. The context window is the problem. Almost everything else Tropo does is the answer.

Beginner Path · New to Agents ← Before: What is an LLM (the model underneath) · Next: What is Memory (how agents keep what the window forgets) → Related: What is a Session · See the whole path: New to Agents - Beginner Path

Pilot/mold article for the beginner path. Diagram is on the Tropo neutral palette; brand-grade illustration is a later designer pass.