Anchoring-A harness for long-term memory maintenance and context window saturation prevention in LLM sessions

Description

An implementation note describing how two previously published components — Cognitive Relay (response-schema field separation) and Memory Spine (hierarchical compressed long-term memory in the assistant role) — are combined into a single harness. It covers how a request is assembled after user input, the structure in which the model responds, where return values are stored, and the path through which memory is compressed and reincluded in the next request. The structure maintains a single logical session without context saturation, using a structured response whose thought field is persisted as plain text rather than as reasoning effort. This document deals only with general examples and their technical elements; it does not include the structure currently under implementation.

Files