Consolidator Enables Persistent Memory for LLMs Across Contexts.

Large language models often struggle with retaining information across distinct conversational turns or context boundaries, effectively resetting their "short-term memory" (STM). While copying STM to a slower "long-term memory" (LTM) store can preserve state, merely retaining it doesn't guarantee that this information will influence future memory access. This paper explores this distinction with Consolidator. Consolidator is a shared, slot-local operator integrated into a Phasor Memory Network (PMNet). It transforms routed STM before accumulating it into LTM, crucially without replaying the original source tokens. After each consolidation, the model's KV cache and STM are cleared, simulating a context boundary. The retained LTM is then not only retrievable but also feeds into a hierarchical router, conditioning which explicit memory slots subsequent inputs access. Evaluated on a two-segment modulo-10 mapping task, Consolidator dramatically improved the recall of updated mappings from LTM, raising it from 44.38% to 87.02%. This significant gain was achieved with only a tiny fraction of trainable parameters (0.041% of the model), demonstrating that learned consolidation, especially when combined with LTM routing, far outperforms simple identity accumulation. This mechanism effectively allows models to learn how to make memory persistent and actively influence future processing.