Identify the failure when persistent memory and fresh retrieval disagree about a fact

When persistent memory says one thing about a fact and a freshly retrieved document says another, what does the model do? The honest answer is: there is **no built-in preference**. The model reads both blocks as text. It does not know which one came from a vector store and which one came from a long-term memory file. It does not have a hard-coded rule like 'retrieval is more authoritative than memory'. What actually drives the outcome is mechanical: **where each block sits in the prompt** (top, middle, bottom), **how recent each block is** to the user turn, and **whether anyone wrote down a precedence rule** the model can follow. The three wrong options all describe behaviors people assume the model has but it does not. The model does not prefer retrieval automatically. It does not pick the shorter source. It does not refuse on conflict (unless you tell it to). It just produces text influenced by what it saw, in the order it saw it. The context-engineering fix is to make precedence explicit. One sentence in the system prompt,'When persistent memory and retrieved documentation conflict, prefer the documentation and note the conflict', removes the ambiguity and stops the outcome from depending on accidental positioning.

**No built-in preference. Outcome depends on position, recency, and explicit precedence rules.** This is the core context-engineering failure mode behind a class of bugs that look like model hallucinations but are actually composition bugs. The team adds a memory layer, the team adds a retrieval layer, and at some point the two layers contradict each other on a fact (the user's company name changed, a feature was renamed, a policy was updated). The system produces the wrong answer and the blame goes to the model. The real cause is that nobody told the model what to do when the two sources disagree. Three levers determine the actual outcome. **Position**: blocks near the top and the bottom of the prompt get more attention than blocks in the middle. If the stale memory is near the bottom and the fresh retrieval is in the middle, the stale memory tends to win. **Recency**: in a streaming conversation, the most recently added block carries weight; if memory was just refreshed but retrieval was prepared earlier in the trajectory, the order matters. **Explicit precedence**: a written rule in the system prompt overrides both effects when the model follows instructions reliably, which modern frontier models do. The right move is to never leave precedence to chance. State the rule explicitly: 'When persistent memory and retrieved documentation conflict on a fact, prefer the documentation, quote the document, and note the conflict to the user.' Now the resolution is deterministic and auditable, not a function of accidental ordering.