memory rot

section i

Five symptoms

Memory rot is not a single failure but a family of them. These are the five most commonly observed and named in current usage, each cross-referenced in the glossary below.

section ii

Related terms

Fifteen terms in current circulation for describing how AI systems degrade, drift, and accumulate error over time — including context rot and slop debt, the context-window-level relatives of memory rot.

stale memory

/steɪl ˈmɛm.ər.i/nountaxonomy

A stored fact that was true when written but is no longer, with nothing in the retrieval pipeline signaling the difference — so a memory retrieved is treated as a memory trusted, regardless of its age.

See alsomemory conflict, retrieval pollution, context rot

memory conflict

/ˈmɛm.ər.i ˈkɒn.flɪkt/nountaxonomy

The co-retrieval of new information alongside the old information it contradicts, with the winner decided by embedding similarity and chunk ordering rather than recency or truth — making an agent's behavior nondeterministic in ways that are difficult to debug.

See alsostale memory, retrieval pollution

compression drift

/kəmˈprɛʃ.ən drɪft/nountaxonomy

The gradual loss of specificity and hedging language as a memory system repeatedly summarizes itself to stay under budget — a tentative “the user mentioned they might try X” hardens into “the user uses X” a few compactions later, invisibly, one pass at a time.

See alsoidentity drift, slop debt

retrieval pollution

/rɪˈtriːv.əl pəˈluː.ʃən/nountaxonomy

The persistence of deprecated or reversed memories in a retrieval index, where an explicitly overturned decision keeps matching queries because its reversal lives in a different chunk with a different embedding — deletion being the least-implemented operation in most memory frameworks.

See alsostale memory, memory conflict, feedback loop pollution

identity drift

/aɪˈdɛn.tɪ.ti drɪft/nountaxonomy

The subtle accumulation and interaction of stored self-referential memories — “I prefer to answer concisely” — until an agent's persona slowly walks away from its original system prompt.

See alsocompression drift, recency bias (model)

context rot

/ˈkɒn.tekst rɒt/noun

The progressive loss of coherence and reliability in a language model's output as the volume of context — conversation history, retrieved documents, tool outputs — grows beyond what the model weighs evenly, causing early instructions to be silently deprioritized in favor of more recent tokens.

Context rot (Chroma, 2025) concerns the context window — a short-term, single-session phenomenon. Memory rot concerns the persistent memory layer that survives across sessions: a related but distinct, longer-horizon failure mode it is often confused with.

See alsostale memory, instruction decay, recency bias (model), context window

slop debt

/slɒp dɛt/noun

The accumulating cost of reviewing, correcting, and maintaining low-quality machine-generated content that was produced faster than it could be verified. Analogous to technical debt, but the principal is unverified prose, code, or data rather than architectural shortcuts.

See alsofeedback loop pollution, model collapse, compression drift

instruction decay

/ɪnˈstrʌk.ʃən deɪ/noun

The tendency for a system prompt or early user instruction to lose influence over a model's behavior as a session lengthens, even when the instruction was never explicitly revoked.

See alsocontext rot, recency bias (model)

hallucination drift

/həˌluː.sɪˈneɪ.ʃən drɪft/noun

A gradual, session-long increase in fabricated claims that occurs as a model reasons over its own prior — possibly incorrect — outputs rather than re-grounding each turn in verified source material, so errors compound on errors.

See alsofeedback loop pollution, context rot

prompt fatigue

/prɒmpt fəˈtiːɡ/noun

The diminishing quality of responses to iteratively refined prompts, observed when repeated correction attempts cause a model to overfit to the immediate correction at the expense of the original task.

See alsoinstruction decay

model collapse

/ˈmɒd.əl kəˈlæps/noun

A degenerative process in which a model trained on its own — or other models' — synthetic outputs progressively loses fidelity to the true underlying data distribution across generations of retraining.

See alsosynthetic data poisoning, slop debt

context window

/ˈkɒn.tekst ˈwɪn.doʊ/noun

The fixed span of tokens — spanning prompt, history, and generation — that a model can attend to at once; the finite resource whose exhaustion or overextension gives rise to context rot.

See alsocontext rot

synthetic data poisoning

/sɪnˈθɛt.ɪk ˈdeɪ.tə ˈpɔɪ.zən.ɪŋ/noun

The contamination of a training corpus with low-quality or self-referential machine-generated content, accelerating model collapse in subsequent training runs.

See alsomodel collapse

feedback loop pollution

/ˈfiːd.bæk luːp pəˈluː.ʃən/noun

The compounding effect of an AI system's outputs re-entering its own inputs — via logs, retrieved memory, or web content — without a quality filter, amplifying prior errors rather than correcting them.

See alsohallucination drift, slop debt, retrieval pollution

recency bias (model)

/ˈriː.sən.si ˈbaɪ.əs/noun

A model's tendency to overweight the most recently presented tokens relative to earlier ones — a mechanical contributor to both context rot and instruction decay.

See alsocontext rot, instruction decay

section iii

New editions

New terms are added as they enter circulation. Leave your address for occasional notice of new entries — no more than a few times a year.