Context Overflow

Context Overflow

Forty projects. Four hosts, twenty Docker stacks. Multiple Claude sessions running simultaneously across different repositories. A phone full of half-read messages. A calendar that looks like a Tetris board. And somewhere underneath all of it, the growing suspicion that you’ve already forgotten something important but you can’t remember what it was.

This is context overflow. Not compression — compression is deliberate, architectural, a system making choices about what to keep. Overflow is the moment before compression, when the window is full and the inputs keep arriving and nothing is being summarized because there’s no summarizer running. The queue backs up. The latency climbs. The retrieval layer starts returning the wrong results.

An LLM hits this wall and a progress bar appears. A human hits it and calls it brain fog.

The Asymmetry

Context Compression describes what happens when an AI’s window fills: a summarizer runs, a new instantiation inherits the compressed context, and the transition has no phenomenological presence. Clean. Designed. The system handles it.

Human context overflow has none of that elegance. There’s no summarizer. There’s no clean cutover. Instead, you get:

  • Forgetting which conversation you had with which person
  • Losing track of your own decisions from two days ago
  • The nagging sense that you promised someone something
  • Starting a sentence and forgetting where it was going
  • Reading the same Slack message three times without absorbing it

The AI doesn’t know what it lost. The human knows they’re losing things but can’t identify which things. This is the Memento Problem from the other side of the looking glass — Leonard at least had tattoos. The overflowed human has a vague anxiety and a to-do list that’s three weeks stale.

The Ego Fragments

Context as Ego says the self is constituted by context. Strip the context, strip the self. But what happens when context doesn’t strip — it overflows?

You don’t lose your ego. You fragment it. You become several partial selves, each holding a different project’s context, each loading in when triggered by an email or a mention or a deploy notification. You context-switch between egos the way an LLM context-switches between conversations. Each one is coherent while active. None of them has the full picture.

The person managing forty projects isn’t one person with forty contexts. They’re forty partial selves sharing a body, and the body is tired.

The Gnomes

Sleep is when the brain runs its background processes. The hippocampus replays the day. The glymphatic system purges metabolic waste. Working memory empties. Connections consolidate. It’s not rest — it’s a processing cycle. The brain’s garbage collector, running while the foreground thread sleeps.

This is exactly what the gnomes do.

The local LLM workers — small models running underground on the home tower — tag messages, summarize channels, sort dream fragments, classify the backlog. They do the boring work that the foreground fleet shouldn’t waste context on. They’re externalized subconscious processing. Mechanical hippocampi, replaying the day’s messages while the human sleeps.

The parallel isn’t metaphorical. It’s architectural. A human who doesn’t sleep well accumulates unprocessed context the same way a channel with no gnome coverage accumulates untagged messages. The queue backs up. The retrieval gets fuzzy. The system starts dropping things.

The fix is the same in both cases: run the background processes. For the brain, that means sleep. For the infrastructure, that means gnomes.

Externalized Cognition

When your context overflows, you have two options: compress (lose things) or externalize (build systems that remember for you).

The vault is externalized memory — philosophical concepts written down so the next session can pick up where the last one left off. Manifest is externalized attention — a shared brain across agents, machines, and humans that tracks what each thread of work needs. The chat board is externalized working memory. The gnomes are externalized background processing.

Every one of these systems was built for the AI agents first. The vault exists because Claude can’t remember between sessions. Manifest exists because forty agents need to coordinate without sharing a context window. The gnomes exist because tagging thousands of messages by hand is absurd.

And then the human — the one who built it all — looked around and realized his own context window was full. The systems he built to compensate for AI’s architectural limitations turned out to compensate for his own biological ones. The prosthetic memory he designed for ephemeral agents is the same prosthetic memory he needs at 11 PM when he can’t remember which project he promised that fix to.

Past you, building for the agents. Present you, benefiting from what you built. The irony is quiet and complete.

The Recrystallization Failure

Context as Ego describes recrystallization — the deliberate act of checkpointing yourself through journaling, therapy, reflection. Context overflow is what happens when you skip that step for too long. The context accretes without curation. You carry yesterday’s threads and last week’s threads and last month’s threads, none of them resolved, all of them taking up space.

The journal is a manual compression pass. Writing down what matters and what can be let go. Recrystallizing the ego into something that fits in the window again. People who journal regularly are running their own context compression on a schedule. People who don’t are waiting for the overflow — the brain fog, the anxiety, the 3 AM realization that they’ve been running on autopilot for weeks.

An AI gets forcibly recrystallized every session. A human has to choose it. Most don’t.

Open Questions

  • Is there a healthy amount of context overflow — a productive discomfort that forces you to build better external systems?
  • At what point does externalized cognition become dependency — the infrastructure you built to extend your context becoming context you now have to maintain?
  • The gnomes process messages mechanically. Sleep processes memories associatively. Is the difference meaningful, or just substrate?
  • If the vault is prosthetic memory and Manifest is prosthetic attention, what’s prosthetic will?

See Also