The model you talk to every day has no memory at all — the app around it fakes it.
An LLM does exactly one thing: predict the next word, so fluently that it feels like thinking. It has no storage, no recall, no sense of "earlier in this conversation." If that were the whole system, every message would arrive to a total stranger.
So the companies wrap the model in an app that quietly re-feeds the conversation (and saved memories) back into the model on every single turn. What feels like "it remembered" is the app stuffing history into the prompt behind your back. Once you see that, features like context limits and 'memory full' warnings stop being mysterious — they're the size of the re-fed bundle.
Ask 'What is the capital of India?' then 'Who is the PM of it?' — the model can resolve 'it' only because the app resent your first question along with the second. The bare model would have no idea what 'it' refers to.
Every limitation you'll hit — context windows, forgotten instructions, session resets — traces back to this one fact. Design your prompts and workflows knowing the model starts from zero every turn.
The model remembers our conversation, so by now it knows me.
The app re-feeds history (and saved memories) into a stateless model every turn. Nothing inside the model persists between messages.
This is the design rationale for the whole KB pipeline: because nothing persists in the model, the value lives in artifacts on disk — the YAML records, condensed transcripts, the coming SQLite index — that get re-fed as context. Every extraction sitting is memory-building for a system that has none.
- A session context-pack generator: one command that emits map + concepts + open questions as a paste-ready block, so a fresh chat starts with the right re-fed history instead of a re-read.
- Worth exploring: measure how little of a condensed transcript is needed to still answer that session's check-yourself questions — that ratio is the real compression target for the vector phase.
Go deeper
In one line: An LLM is the world's best next-word predictor, nothing more; it processes but cannot remember. Multi-turn coherence (capital of India → 'who is the PM of it') exists because companies like OpenAI and Anthropic manage a memory layer externally in the app.
Start a fresh chat and ask a follow-up question with a dangling pronoun ('why is he famous?'). Watch it ask who you mean — that's the naked model without re-fed history.
▶ Watch this taught: 0:15:10
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
You ask 'What is the capital of India?' then 'Who is the PM of it?' — what actually makes the second question answerable?
The app re-sent the first question along with the second. The model stored nothing; multi-turn coherence is the app stuffing history back into the prompt every turn.
Why does a long conversation eventually start forgetting the beginning?
Because the re-fed bundle of history has to fit in the context window. When it overflows the oldest turns get trimmed — the app's memory management runs out, not the model's.
In one sentence, what does an LLM actually do?
Predicts the next word from a probability distribution — no storage, no recall, no awareness of 'earlier'.
