BLOG
Google Research separates fast context, low-rank expert consolidation, and synthetic dreaming to improve continual learning without simply enlarging the active context window.
Anthropic lays out a four-layer agent-memory hierarchy and shows how dreaming improves recall, corrects errors, and builds navigable knowledge.
MemCurrent pairs append-only episodic traces with a distilled file-style knowledge base, wired in as plugin, tools, and offline dreaming. It triples pass rate on an internal benchmark (22.3% to 68.4%) with 31% fewer tokens.
Tencent Hunyuan open-sources Hy Memory 2.0: the 3B-active writer model Hy-Memory-30B-A3B, the hy-memory SDK, and local-first MemHub; a 55.49 macro average over four harnesses edges out Opus 5.
As an Amazon Associate, we earn from qualifying purchases.
