Retrieval failures: getting the wrong memories back

The "Practical Guide" details several ways the read phase goes wrong:

• Semantic versus causal mismatch. Similarity search returns memories that look related but are not. Embeddings are good at "this text looks like that text" but bad at "this is the cause of that." The author sees this when debugging through coding assistants: they spot a similar-looking error and miss the real underlying cause, leading to thrashing, many changes that never fix the actual problem.
• Memory blindness. In tiered systems, an important fact never resurfaces. The data exists, but the agent never sees it again, perhaps because a sliding window moved past it, or because the system retrieves only the top 10 memories and the one it needed was the 11th.
• Silent orchestration failures. The author calls these the most dangerous. The paging, eviction, or archival logic does the wrong thing, but no error is thrown. The only symptom is that responses slowly get worse, more generic, and less grounded. His real example: his system silently failed to write daily memory files, so the daily summaries had nothing to work from, and he only noticed because the agent kept forgetting recent work.