Search papers, labs, and topics across Lattice.
The paper introduces Memini, an external memory system for LLMs inspired by biological memory, featuring coupled fast and slow variables on graph edges to model synaptic consolidation. This design enables episodic sensitivity, gradual consolidation, and selective forgetting within a unified framework, allowing the external memory to adapt and reorganize knowledge through its own dynamics. The authors argue that this approach is more biologically plausible and effective than explicit memory management in LLMs.
Forget rigid memory structures: Memini lets your LLM's external knowledge evolve organically, learning and forgetting like a brain.
LLMs are trained once, then deployed into a world that never stops changing. External memory compensates for this, but most systems manage it explicitly rather than letting it adapt on its own. Biological memory works differently: coupled multi-timescale dynamics make new associations immediately usable, strengthen what repetition confirms, and let the rest fade. We argue that external memory should follow a similar principle. In Memini, this view takes the form of an associative memory that organizes knowledge as a directed graph. Each edge carries two coupled internal variables, one fast and one slow, following the Benna-Fusi model of synaptic consolidation. From this coupling, episodic sensitivity, gradual consolidation, and selective forgetting emerge as facets of a single mechanism, reframing external memory as a learning substrate that reorganizes through its own dynamics.
