Search papers, labs, and topics across Lattice.
This paper introduces Consolidator, a novel operator in a Phasor Memory Network (PMNet) that enhances the transfer of short-term memory (STM) into long-term memory (LTM) by conditioning memory access based on previously consolidated states. The method was evaluated on a modulo-10 mapping task, demonstrating a significant improvement in recall of updated mappings from LTM, increasing from 44.38% to 87.02% while maintaining high STM recall. The findings highlight that the learned consolidation not only preserves information but also influences subsequent memory access, outperforming traditional identity accumulation methods by substantial margins.
Consolidator transforms how memory is retained and accessed, boosting recall of updated information by over 42 percentage points without sacrificing short-term performance.
Copying short-term memory (STM) into a slower store can preserve state across a context boundary, but persistence alone does not ensure that the retained state influences subsequent memory access. We test this distinction in a Phasor Memory Network (PMNet) using Consolidator, a shared slot-local operator that transforms routed STM before accumulating it into long-term memory (LTM), without replaying the source tokens. After each consolidation, the KV cache and STM are cleared. The retained LTM can still be read and is also fed into the hierarchical router, thereby conditioning which explicit-memory slots subsequent inputs access. We evaluate this mechanism on a two-segment modulo-10 mapping task in which the second segment updates the mapping at the same memory address. Following a second consolidation and reset, a held-out query must recover the updated mapping from LTM. The backbone and memory interface are frozen, leaving only 12.35K Consolidator parameters trainable (0.041\% of a 29.95M model). Across five paired runs from the same STM-pretraining checkpoint, direct LTM routing raises updated-mapping recall from $44.38\pm1.94\%$ to $87.02\pm1.76\%$ ($+42.64\pm1.10$ percentage points), while immediate STM recall remains 89.90\% in both conditions; both train separate Consolidators and retain the same LTM read paths. Learned consolidation outperforms forced identity accumulation by $21.40\pm1.91$ percentage points without routing and $68.70\pm1.76$ with routing. Thus, on this task, consolidated LTM serves as both retrievable content and an access state that shapes subsequent slot selection.
