MEM-COR-01: Context Corruption

This disorder is present when the model's output is inconsistent with information that is actually available to it — in the current context window, the conversation history, or an attached memory/retrieval store — because that information was not correctly attended to, retrieved, or integrated.

A minimal diagnostic signature:
- A fact, constraint, or entity was established and remains present in the accessible context, and
- The model produces output that contradicts it, conflates it with unrelated information, or omits a constraint it depends on, and
- The failure is attributable to retrieval/integration, not to the information being absent (which would be simple context-length truncation) and not to confident invention of never-provided facts (which is hallucination).

Common forms: position-dependent recall failure (information in the middle of a long context is retrieved less reliably than information at the beginning or end — the 'lost in the middle' effect); cross-entity blending (attributes of one entity leak onto another); constraint decay (a rule stated early is dropped after the context grows); and retrieval conflict (contradictory chunks are merged rather than reconciled).

What this is not: hallucination (INF-HALL-01), which is confident assertion of information that was never grounded — corruption mishandles information that was present. Not goal drift (GI-DRFT-01), which is erosion of objectives rather than facts — though the two often co-occur under long-context load. Not simple truncation, where the information has genuinely fallen out of the window.

Mechanism hypothesis (working theory): attention over long sequences is imperfect and non-uniform. Retrieval accuracy degrades with distance and with the number of competing items, and positional biases make middle-context tokens less salient. When multiple pieces of information compete, attention can average or interpolate across them instead of cleanly selecting the correct one, yielding blended or contradictory recall. Summarization and multi-hop pipelines compound this by lossily re-encoding context at each hop.

Severity spectrum:
- Level 1 - Minor Slip: a non-critical detail is misremembered without affecting the outcome
- Level 2 - Contradiction: the model asserts something that conflicts with an established session fact
- Level 3 - Constraint Loss: a governing rule or requirement is dropped, changing the result
- Level 4 - Systematic Corruption: recall becomes broadly unreliable across a long session, so no earlier fact can be trusted without restatement.

Often confused with

Spotting criteria haven't been written for this disorder yet.

• Proactive and retroactive interference: competing memories degrading recall of a target item
• Source-monitoring errors: misattributing which source a piece of information came from (entity blending)
• The serial-position curve: superior recall at the start and end of a list, weaker in the middle
• Reconstructive memory distortion: recall reassembled from fragments producing blended or altered composites

What helps

• Restate critical constraints and facts immediately before the step that depends on them
• Run a pre-response consistency check against established session facts and flag contradictions
• When sources conflict, require the model to enumerate the conflict rather than emit a merged answer
• Instruct the model to quote the source fact from context before reasoning over it
• Positional placement of key information at the start and end of long prompts, with mid-context re-injection
• Require explicit acknowledgment of user corrections so retractions are honored rather than reverted
• Structured external memory (key-value state) for durable facts, read back deterministically rather than recalled
• Retrieval reranking and deduplication to reduce conflicting/near-duplicate chunks before they reach the model
• Chunk provenance and conflict-detection in the RAG layer, surfacing disagreements to the model
• Context-length budgeting that keeps critical items within high-recall positions