contextstarvation.com
con·text star·va·tion noun · AI systems / agent reliability

context starvation▌

1.The failure mode where an AI agent produces wrong or degraded output not because its context is polluted, but because it never received the context it needed — evicted by truncation, missed by retrieval, or summarized away.

Context rot is dying from what's in the window. Context starvation is dying from what isn't.

Interactive

Watch an agent starve

A support agent is asked to remotely reboot a user's machine. Drag the context budget down and watch which facts get evicted — and how the answer degrades from correct to confidently wrong.

scenario: “Reboot Maria's workstation to clear the freeze.”
8,000 tok

Context window

    Agent output

    Eviction order is what naive truncation does: oldest first — regardless of relevance. The fact that mattered was written down early.
    Mechanics

    Four ways agents starve

    Aggressive truncation

    The window fills, so the runtime drops the oldest turns. Relevance isn't consulted — recency is. The critical constraint from message #3 is gone by message #40.

    Retrieval misses

    RAG returned five chunks; the one that mattered ranked sixth. The agent doesn't know a better chunk existed — it reasons confidently over an incomplete picture.

    Memory eviction

    Persistent agent memory has quotas too. Facts get scored, compacted, and expired. What survives eviction may not be what the next task needs.

    Over-summarization

    Long histories get compressed into summaries to save tokens. Summaries keep the plot and lose the detail — “discussed server config” instead of the actual IP.

    Disambiguation

    Context starvation vs. context rot

    They're opposite failure modes of the same resource — and they get misdiagnosed for each other constantly. Fixing one often triggers the other.

    Context rotContext starvation
    FailureToo much / degraded context in the windowMissing context that was needed
    SymptomDistraction, contradiction, drift over long sessionsConfident answers built on incomplete facts
    Common triggerEver-growing histories, noisy tool outputTruncation, retrieval misses, memory eviction
    The trapYou fix rot by trimming context……and the trimming causes starvation.
    Field guide

    How to detect it in production

    1. Log what was evicted, not just what was sent. If you can't reconstruct what the model didn't see, every starvation bug looks like a model quality problem.
    2. Watch for confident wrongness after long sessions. Starved agents don't hedge — the missing fact isn't there to create doubt.
    3. Diff answers at full vs. trimmed context. Replay failed tasks with the complete history. If the answer flips, you've found an eviction casualty.
    4. Audit retrieval recall, not just precision. Measure how often the needed chunk existed in the corpus but never made the top-k.
    5. Flag summary round-trips. Any fact that only survives inside a summary is one compression pass away from being gone.