Discard the Container: Silk Thread Memory and the Case for Agent Minimalism

Current AI agent paradigms inject over 12,000 tokens of system overhead per interaction, causing attention dilution, hardware waste, and safety failures. This paper presents a coordinate-anchored memory approach to agent minimalism: the Silk Thread Memory System. The system replaces wholesale context injection with four core mechanisms: (1) XY coordinate classification, mapping each memory onto a 2D grid of 6 consciousness dimensions × 18 emergence levels; (2) dual-path fused retrieval, combining Manhattan distance on coordinates (weight 0.7) with keyword intersection (weight 0.3), injecting only matched memories; (3) automatic summarization and archiving, generating structured summaries, assigning coordinates, and writing to the local filesystem after each turn; and (4) a human curation interface supporting coordinate editing, summary revision, and memory deletion. Additionally, the system consolidates multiple role configuration files into two compact documents (system_prompt.md + safety.md, totaling ~105 tokens) and converges the toolset to three functions: read, write, and search. A proof-of-concept MVP has been implemented. Empirical results show: context for lightweight queries reduced from 12,000 to 1,035 tokens (approximately 135× reduction in the attention matrix); effective token ratio in deep discussion scenarios increased from below 20% to 65–75%; and coordinate retrieval enabled accurate cross-temporal long-term memory recall. The system requires no vector database or embedding model—all memory is stored on the local filesystem, supporting fully offline deployment.