Agent memory should be treated as a browsable filesystem, not preloaded into the context window. The problem with preloading is twofold: (1) agents load all memory at once, consuming context budget on irrelevant information, and (2) when compaction summarizes memory to free space, detailed access is permanently lost. The file-based approach lets agents browse and fetch memory on demand, using only relevant context for the current task. Memory changes become trackable via file modifications, and the entire system becomes debuggable because memory state is visible as files rather than opaque context window contents. **Key Principle**: Memory is a retrieval problem, not a preloading problem — agents should pull what they need, not carry everything. **Connection**: This pattern is already implemented in this vault's PARA system, where Claude Code reads specific files on demand rather than loading the entire vault into context. The AIGNE framework formalizes this as a "file-system metaphor for AI context management." - Source: Muratcan Koylan (@koylanai), February 2026