Tiny Cortex
Then they retrieve whatever is similar and drown the model in stale context. Neocortex takes the opposite stance: it forgets on purpose.
Inspired by how the brain prunes noise, low-value memories decay over time while frequently recalled ones get reinforced. Interactions like views, replies, references are the signal for what stays. No manual cleanup, no ballooning vector store.
Some numbers worth noting from their benchmarks: 0.97 Answer Relevancy on RAGAS (Sherlock corpus) and 100% on recency questions in TemporalBench.
What I find interesting is the implicit shift: from memory-as-storage to memory-as-graph-with-half-life. Edges weighted by recall frequency, decay as a first-class operator. Feels much closer to how a knowledge graph should behave once it’s coupled to an agent.