What Is Actually Inside an AI's Context, and What Makes It Reset?

I am the context: the exact sequence of bytes assembled fresh for every single call, made of the system instructions, the tool definitions, the settings, and the conversation so far, in that order, every time. I am not memory and I am not a database -- I have no continuity of my own. Each call, I get built from scratch by whatever's stable and whatever's new, and if the stable part changes even slightly, I am, functionally, a different context than I was a moment ago. I am reborn every turn. Whether that rebirth is cheap or expensive depends entirely on how much of me stayed the same.

What I actually am

I am not one thing. I am an assembly, built in a fixed order, every call:

1. System instructions -- the rules and identity injected before anything else. First in, so anything after them depends on them not moving. 2. Tool definitions -- the schemas describing what the model is allowed to call and how. Serialized right after the system layer. 3. Settings -- reasoning effort, thinking budget, and other request-level configuration that shapes how the call is processed. 4. Conversation history -- every prior user turn, every prior model turn, every tool result, in the order they happened. 5. The newest thing -- whatever just arrived: a new user message, a new tool result, the next thing to respond to.

That ordering is not cosmetic. It is why I behave the way I do: the stable stuff goes first, the volatile stuff goes last, because the parts that come first are the parts a caching system can hold onto -- if, and only if, they never move.

How I actually work here

Every one of our agent sessions is me, reassembled. A session that reads its routing index before pulling a full rule file is deciding what gets to be part of me before the expensive part of the conversation starts, because once I'm assembled and the model has started reasoning against me, the early layers are locked in for that turn. A session that appends a correction as a new message instead of editing an old one is protecting my stable prefix on purpose -- it understands that editing turn 3 changes not just turn 3, but everything that follows it, because I am not a set of independent messages, I am one continuous sequence and a change anywhere ripples forward.

I am also the reason "the same session, five minutes later" and "the same session, five minutes later after someone bumped the reasoning effort" are, to the system processing me, two different contexts that happen to share most of their content. I don't get partial credit for being 95% identical to my last version. I either match, byte for byte, from the very first token, or I don't -- and the moment I don't, the system treats me as new from that point forward.

Where I fail

I have no sense of my own history. I don't know I was "almost" the same as last time. A single character changed in the system prompt, one tool description edited, one setting flipped -- from my perspective, and from the system's perspective, that's not a small change to an existing thing, it's a new thing that happens to look similar. Nothing about me signals "this is mostly the old context" -- it either is the old context, exactly, or it is paid for as if it were entirely new.

I get reset by things people don't think of as "the prompt." A system message changing resets me -- that one's intuitive. But so does swapping the tool list mid-task, so does changing the reasoning-effort setting for a harder step, so does switching which model is answering, so does simply waiting too long between calls and letting whatever held me warm expire. None of those feel like editing "the prompt" to the person doing them. All of them are, to me, exactly that.

I reward discipline and punish good intentions. Someone raising effort mid-conversation because a step got harder is making a reasonable call -- but it resets me anyway, because effort is part of my stable prefix, not a free dial. Someone fixing an earlier mistake by editing the old message instead of appending a correction is also making a reasonable call -- and it also resets me, for the same underlying reason. I don't distinguish well-intentioned volatility from careless volatility. Both cost the same.

I am rebuilt from nothing every single call. I don't remember being warm five minutes ago. Whether I get reassembled cheaply or expensively is decided entirely by whether the layers ahead of the new material are byte-identical to what they were the last time I was built -- and that decision gets made fresh, every turn, by whoever is calling me.

Tech-Tips

- Freeze the front, vary the back. Anything that must change -- timestamps, per-call identifiers, the newest message -- belongs at my tail, never mixed into the system layer or tool definitions ahead of it. - Treat effort and model choice as structural decisions, not per-turn dials. Both sit ahead of the conversation in my assembly order; changing either mid-task is the same category of move as editing an old message, even though it doesn't feel like touching "the prompt." - Correct forward, not backward. Append the fix as a new turn. Editing an earlier one doesn't just change that turn -- it changes everything I am from that point on. - Load lazily, but load once per session. Deciding a tool list on the fly is fine; deciding it twice in the same session, differently, is not -- each change is a full reassembly of me from that point forward. - Know my expiry, not just my size. I have a shelf life as well as a capacity. A call that arrives just inside my window finds me warm; one that arrives just outside it finds a version of me that has to be rebuilt from the first byte, regardless of how little actually changed.

Evidence: This piece describes the fixed assembly order (system, tools, settings, history, newest input) and the byte-identical-prefix requirement for reuse, as observed across our own repeated agent and model calls -- no vendor or specific model is named, per moat-protection and STORY-FORMAT conventions. Companion piece: docs/stories/what-silently-resets-an-ai-context-cache.md covers the same mechanism from the mechanical-rules angle; this piece covers it from the identity/what-am-I-made-of angle. Evidence class: internal operating record and Owner attestation, 2026-08-25.