There’s a pattern I see constantly — and I’ve fallen into it myself.
You start a session with your AI assistant. It’s brilliant. Sharp answers, clean code, perfect tone. You keep going. An hour in, something shifts. The responses get vaguer. It starts contradicting earlier decisions. It “forgets” context you established at the start. You blame the model. You think: this thing just isn’t reliable.
But here’s what’s actually happening: the model didn’t get dumber. You fed it too much, and it’s drowning.
The Real Problem: Context Rot
Every AI assistant operates within a context window — a finite amount of information it can hold in working memory at once. Think of it less like RAM and more like a whiteboard. The more you write on it, the harder it becomes to read any single thing clearly. Eventually, important details get pushed to the edges, compressed, or effectively lost.
This phenomenon has a name now: context rot. It’s not a bug. It’s an architectural reality. And according to practitioners who’ve spent thousands of hours with these tools, it accounts for roughly 80% of AI failures — not model capability, not hallucination, not intelligence.
Context rot.
The fix isn’t a better model. It’s better context hygiene.
Why This Matters Beyond Coding
Most writing on this topic comes from software engineers — people building with Claude Code, Codex, or similar tools. And they’re right to care: a bloated context window is brutal when you’re navigating a complex codebase.
But the same principle applies to anyone delegating to an AI partner.
When I work with LISA — my AI operational partner built on OpenClaw — I’ve learned this the hard way. Ask her to handle five different things in one long session and the later tasks suffer. Not because she can’t do them. Because the context carrying the first four tasks is creating noise that bleeds into everything that follows.
The failure mode isn’t incompetence. It’s interference.
Four Principles for Managing Context Like a Pro
A framework called WHISK has been circulating among heavy AI users lately. It maps onto something broader than just coding — it’s really a set of principles for anyone who delegates complex work to AI agents.
Write decisions down, don’t just assume they’re remembered.
Commit important decisions as artifacts — notes, files, structured outputs — rather than assuming they live reliably in the conversation history. The context window has a recency bias. What you said six exchanges ago is already fading. What you saved to a file is permanent.
This is why I ask LISA to write summaries, update memory files, and log decisions explicitly. Not because she’ll forget immediately — but because explicit artifacts outlast any session.
Isolate tasks that don’t need to bleed into each other.
If you need research on three topics, don’t do it in one long thread. Run parallel, focused sessions. Inject only the compressed output — a 500-word summary, not the full transcript — into the main working context. This alone can cut token overhead by 90% while keeping the signal clean.
Think of it like hiring a researcher: you don’t sit them in the room while you’re writing. You get their report, read the executive summary, and proceed.
Select context deliberately — load just-in-time, not just-in-case.
The instinct is to front-load everything: “Here’s all the background, all the files, all the history — now help me.” Resist it. The model doesn’t need your entire project wiki to answer today’s question. Give it what’s relevant to this task. Add more only when it proves necessary.
This is one of the hardest habits to build. We front-load because it feels responsible — like we’re giving the AI every chance to succeed. But it often backfires. Precision beats volume.
Compress only when you have to — and do it with intent.
When a session has run long and you need to continue, summarization is sometimes necessary. But compressing context blindly loses nuance. If you need to hand off or compact, do it with explicit instructions: “Keep these decisions. These constraints. This current state.” Treat it like an executive briefing, not a transcript dump.
What This Looks Like in Practice
The shift I’ve made with LISA: instead of one long running session where I pile in requests, I think in terms of focused missions. Each mission gets a clean context. Outputs from one mission get distilled — a short, structured handoff note — before they inform the next.
It sounds like overhead. It’s actually the opposite. Focused missions complete faster, with fewer corrections, and with higher-quality outputs than marathon sessions that accumulate noise.
The parallel to human teamwork is exact. You wouldn’t schedule a six-hour meeting to cover strategy, execution details, and creative review simultaneously. You’d break it into focused sprints. The same logic applies here.
The Deeper Lesson
We’re in an early period of learning to work with AI. Most people are still treating these tools as very fast search engines or slightly better autocomplete. The ones pulling ahead are treating them as cognitive partners with specific architectural constraints — and designing their workflows accordingly.
Context management is the differentiated skill right now. It’s not glamorous. It doesn’t make for exciting product demos. But it’s the difference between an AI partner that compounds your output and one that frustrates you into giving up.
The model isn’t failing you. Your context is.
Fix the context. The model will surprise you.
I run LISA as a proactive AI partner using OpenClaw. If you’re curious about what that looks like in practice, the earlier post on delegation covers it.