The model is a given.
The context is the craft.
A practical tour of context engineering. One character runs through the whole talk: a brilliant new hire with no memory.
Same model.
Two answers.
You've hired a genius with no memory.
Brilliant in general. Her name is Leila. But every morning she walks in knowing nothing about you, your work, or what happened yesterday. That's the model. The whole job is managing her well.
The docs, the memory, the tools, the history: everything loaded in front of the model when it answers. Prompt engineering is now a subset.
CPUthe model · fixed, powerful
RAMthe context window · small, precious
OScontext engineering · decides what loads, and when
The model is the CPU.
The context window is the RAM.
Karpathy's framing, and it's the one that sticks. The processor is fixed. The working memory is small and precious. Context engineering is the operating system: it decides what to load into that memory, and when.
When AI fails, it's usually not too dumb.
It just wasn't told what it needed to know. As models get smarter, the failures stop being about the model and start being about the context. So the skill that lasts, for users and builders alike, is context.
More context can make it worse.
Tip every document onto her desk on day one and she drowns. So does the model. Chroma tested 18 frontier models and every one degraded as the input grew, before the window was even full. A bigger desk is not the answer. A tidier one is.
Models forget the middle.
Put the key fact at the start or the end and the model finds it. Bury it in the middle of a long context and accuracy drops 30% or more. And the bigger the window gets, the deeper the sag.
Keep a big freezer. Only cook what's new.
Five levers. One stack. Not rivals.
RAG
What to fetch. Keep bookmarks, not the whole library.
Memory / notes
What to remember. Write it down, reload later.
Compaction
What to forget. Summarise and move on.
Tools
What it can do. The myth-buster is next.
Sub-agents
Divide and conquer. Anthropic's multi-agent setup beat a single agent by 90.2%, and token usage alone explained 80% of the difference.
The model never uses a tool. It just asks.
Tools feel like the model "doing things". They're not. The model only asks. Your system makes the call and feeds the result back into the context. A tool is really a context-builder.
Every so often: summarise, and start fresh.
Claude Code auto-compacts at 95% of the window. In plain ChatGPT or Claude you get no view of how full it is, so be the harness yourself: once a chat gets long, ask it to summarise, copy that, paste it into a fresh chat. It is lossy. The big decisions survive, the buried specifics do not.
There's a right way and a wrong way to forget.
Drop the oldest tokens
Cheap. But it bins the start of the context, and we just saw the model reads the start best.
Keep the edges, summarise the middle
The direct fix for lost in the middle. Protect the edges, compress the sag.
Clear the noise
Strip the old tool calls and results the model no longer needs, keep the reasoning. You trim the context without losing the thread.
A copy of a copy of a copy.
Compact from the source, not by re-compressing your own compressions.
The fix for amnesia is a notebook that compounds.
Give the new hire a notebook they keep and re-read, one that gets richer every day instead of being rewritten from scratch.
A big refactor. And the further it went, the worse it got.
A while back I was running a large refactor across a big codebase. Quality held early, then drifted. The agent started repeating mistakes and losing the thread. That's context rot. Live, on a real job.
So I built a loop that never lets the context rot.
Chunk
The plan into phases, each with clear acceptance criteria.
Fresh context
A harness spawns a brand-new session per phase, so no single window rots.
Exit gate
It only signals done when criteria verify, otherwise it retriggers fresh. Circuit breakers stop it spinning.
Distil findings
A lighter model harvests verified patterns into a compounding skillbook, promoted only after seen twice, and feeds later phases just the relevant slice.
I'd hand-built the entire textbook. By accident.
I built every lever because the quality was rotting and I had no choice. Months later, Anthropic published the same playbook. This isn't theory. It's what you reach for when it breaks.
I had to build the operating system. For you, it's already in the kernel.
Auto-compaction, sub-agents, a memory tool, built-in context management. It ships now. The one move left is to switch it on.
Be honest about the limits.
But the wiki cost 21× more per question and ran 6.6× slower. Small sample, directional.
- Contamination. A wrong fact, once written, gets read as truth by everything after. My "seen twice" gate helps. It isn't free.
- Scale wall. Lovely at 50 to 100 sources. At thousands it's a knowledge base with no editors.
- Vibe thinking. The understanding forms while you do the filing. Outsource that and you get a tidy wiki you never learned.
Waiting for one missing piece.
Every version died on the same rock: maintenance is a chore humans abandon. The model is the first tireless librarian.
I didn't just talk about the five levers tonight. I used every one of them to build this.
The panel of twenty that reviewed this talk. Eighteen flagged the same problem. They were right, and what you just watched is the fix.
- Fetch. Research agents read the 2026 papers, so each slide only carried the facts it needed.
- Remember. What I learned went into notes and a notebook the next session could read. I never argued the same point twice.
- Forget. When a session filled up it wrote a handover note and started fresh. The morning handover from the war story, on this deck.
- Do. The skills did the real work. One narrated this whole script. One drew every photo of Leila, the same face every time, on purpose.
- Divide. 53 agents reviewed it, because I didn't trust my own opinion on its own.
I argued with a focus group of twenty about my own talk, and I lost. Very 2026.
Every one of them did the filing. Not one could tell me what the talk was about, which joke to keep, or that she should be Leila, a person with a name. That part stayed mine.
The machine can do all the filing. The one job it can't do for you is decide what it means. And that was always the point.
Context is the craft
Try it tonight.
Anyone
Before a big ask, give the context first. Paste the brief, the goal, the constraints. And start a fresh chat when it drifts. Watch the answer change.
Hands-on
Point Claude Code or Codex at one folder, ingest five things into a markdown wiki, ask one question. That's a wiki, and you didn't have to build a Ralph loop.
Thank you.
The talk
The model is a given. The context is the craft.
A practical tour of context engineering.
Prepared by
Amir Zahedi · fleetguru.ai
Sources
Anthropic, context engineering (Sep 2025) · Chroma, context rot · Karpathy · Liu et al., lost in the middle · Lütke (qmd) · Bush, Memex · Luhmann, Zettelkasten · Forte, Second Brain · LOCA-bench (2026) · Lost in the Noise (2026) · Wu et al., position bias (MIT, ICML 2025) · Anthropic, server-side compaction + Memory for Managed Agents (2026).