Generation-free hierarchical KV-cache memory, foveated context navigation, and continual memory harnesses for long-horizon AI agents.
Baroque Technologies develops inference-time memory architectures for Large Language Models (LLMs) and autonomous software agents. As reasoning and coding workflows scale to ratusan of thousands of tokens across multi-turn sessions, standard full-context attention becomes bottlenecked by accelerator memory and distractor interference, while destructive text summarization and static token eviction lose critical multi-hop state.
Our research focuses on lossless, off-accelerator KV-cache tiering paired with compact, multi-tier relational indices and query-adaptive foveal navigation—enabling frozen and fine-tuned language models to retrieve exact historical spans on demand with bounded active attention budgets.
Generation-free ingest indexing across lexical, conceptual, and emergent entity-level abstractions with lossless host-memory KV storage.
Per-head query-driven context extraction and iterative attention-residual refinement for multi-hop reasoning and variable tracking.
Persistent, zero-re-prefill working memory and automated benchmarking across official RULER, NoLiMa, and LongBench v2 evaluation suites.
Location: Toronto, Ontario, Canada
Research & General Inquiries: research@baroquetechnologies.org