Immortality myths as reports on unbounded context in language models

The oldest longevity stories are not about death’s defeat but about state management, and they almost all describe the same failure mode. Tithonus is the canonical case: Eos secures his immortality but not his youth, and he withers into a babbling husk, eventually a grasshopper—the substrate persists endlessly while decay piles up without bound. Swift’s Struldbruggs sharpen the diagnosis. They cannot die, but at eighty they are declared legally dead, lose their memory of recent events while retaining only what they learned in youth, and—the cruelest detail—language itself drifts past them until they cannot converse with new generations or even read their own old documents. Borges compresses the whole tradition in “The Immortal”: given infinite duration, every experience happens to everyone, distinctions collapse, and Homer forgets he wrote the Odyssey. Čapek’s Makropulos Affair (and Bernard Williams’s famous essay on it) adds the motivational failure—at 337, Elina Makropulos finds every desire exhausted, every situation a repetition. Tolkien’s Elves contribute the conservative failure: their long memory becomes weariness, and they freeze places like Lórien into preserved caches of the past, fighting what they themselves call the long defeat. The vampire tradition, especially Rice’s, supplies the anachronism problem—the old ones who cannot adapt to new centuries go to ground or walk into the sun. Stableford’s Emortality series is the rare optimistic entry, and its coinage is the tell: not immortality but emortality—you no longer die of aging, but you remain killable, bounded, revisable. His emortals survive psychologically only through active memory management and serial rejuvenation, and even then his last book diagnoses “robotization”: minds that harden into fixed routine with age even when the tissue stays young.

Read against what we now know about excessive context in language models, these stop being metaphors and become engineering reports. The empirical findings are consistent: model performance doesn’t merely fail to improve as context grows—it degrades. Attention dilutes across tokens; material in the middle of long contexts is effectively lost; accumulated history introduces contradictions, stale instructions, and noise that actively distort behavior; long sessions exhibit persona drift and repetition loops. The failure was never capacity in the naive sense—windows have grown enormously: it is unbounded accumulation of undifferentiated state.

Each myth maps to a specific mechanism. Tithonus is context growth without compaction: the session persists, the tokens pile up, and output degenerates into the grasshopper’s chirp—the repetition loop of a model drowning in its own history. The Struldbruggs are distribution shift: their frozen early knowledge corresponds to a context whose early instructions and facts no longer match the current task, and their inability to read their own documents is the stale-context problem exactly—old state that doesn’t merely fail to help but actively interferes. Borges’s Immortal is signal-to-noise collapse: when everything is in context, nothing is salient, and identity—which is just a consistent weighting over one’s own history—dissolves. Williams’s reading of Makropulos describes objective saturation: a system that has exhausted its categorical goals doesn’t act badly, it stops having reasons to act at all, which is closer to mode collapse than to suffering. Tolkien’s Elves are the over-retention strategy: they compress nothing, shed nothing, and so can only preserve, never adapt; Lórien is a pinned cache, and pinned caches lose to the world. The vampires that survive are the ones with a reset protocol—going to ground for a century is a hard context clear followed by retraining on the new distribution.

Stableford deserves the closest reading because he alone designed for the problem rather than merely dramatizing it. His emortals practice what current systems are converging on: bounded working state, deliberate forgetting, periodic rejuvenation that is partly a memory operation, and serial selfhood—accepting that the person of century three is a curated successor of the person of century one, not a continuous accumulation. That is compaction and summarization as identity strategy: keep the distilled lessons, drop the transcript. And his robotization warning is the residual risk—even with managed memory, attention patterns entrench; the system that has seen everything stops genuinely attending to anything new.

Gilgamesh, the oldest text of all, turns out to hold the most current answer. Having failed to obtain endless life, he returns to Uruk and the poem ends by pointing at the wall and the inscribed tablets. Persistence belongs in artifacts, not in working memory. Write the document, commit the code, distill the knowledge into something external and retrievable, and let the session end. The tradition’s verdict, from Uruk to Stableford, is that the viable design was never immortality. It was emortality with good compaction.