Tag
The author describes a backend agent memory system that outputs embedded state parameters instead of natural language chunks, using a GGUF adaptor with a qwen3 model. It employs dual perpendicular graphs and cyphers for real-time memory integration, potentially eliminating the need for context windows, RAG, and tool schemas.
MIT CSAIL researchers propose a novel approach to AI memory that avoids context rot by storing documents externally and having the AI navigate and query them, achieving 10 million token effective context at lower cost.
Arkon is a new tool that integrates with Claude via MCP to provide centralized knowledge management and granular access control for organizations, replacing manual context-pasting with an LLM-generated wiki layer.
The paper introduces WiCER, an iterative algorithm for compiling domain knowledge into LLM Wiki systems to minimize information loss and catastrophic failure rates during knowledge distillation. It demonstrates that this approach improves upon full-context KV cache inference by preserving critical facts better than blind compilation methods.
SDSR proposes lightweight self-describing structured data with dual-layer guidance to exploit LLM primacy bias, achieving 100% routing accuracy without vector DBs.
Hindsight is an agent memory system designed to enhance AI agents with long-term learning capabilities, achieving state-of-the-art accuracy on memory benchmarks.