PersMem Makes Agent Memory Personality-Driven

PersMem targets a real gap in persona engineering: agents may speak in character while remembering selectively in ways that contradict it. The approach is promising, though its personality mappings remain designed assumptions rather than validated models of human cognition.

• –Four-way attachment classification reached 48.1%, versus a 25% chance baseline.
• –Big Five dialogue identification reached 67.5%, beating uniformly sampled memory by 6.7 percentage points.
• –On CoSER, PersMem scored 69.33 for character fidelity and 84.33 for storyline quality.
• –Inspectable retention and retrieval traces could make persona consistency easier to debug and evaluate.
• –Longer interactions and human judgments are still needed to establish whether the gains generalize beyond controlled experiments.