Personal queries ask 'What did I discuss with Marco at Tuesday's meeting?' A large language model (LLM) must first find that memory. Standard embedding models, which turn text into numeric meaning, are trained for topic overlap. Changing date, place, person, or event barely changes the embedding, so retrieval fails.
Diagnostics showed missing signals live inside the model, not in its output. The fix adds four small internal adapters, one per axis: time, place, person, event, selected per query. Training touches a small part of the model.
On standard embedding models and personal-memory benchmarks, mean retrieval accuracy improved by about one and a half percentage points, with largest gains on hard multi-evidence questions. This makes AI memory retrieval more sensitive to personal experience.