dsh-memory-loader

DSH natively auto-loads only instruction file chains (AGENTS.md, CLAUDE.md); memory directories for global and project notes rely entirely on the model reading them voluntarily, which is probabilistic and degrades as context grows. dsh-memory-loader turns this into a deterministic, harness-level injection so that every new session reliably receives its two-level memory without hoping the model remembers to do so.

The plugin hooks into the agent/pre-step seam and injects four files in broad-to-specific order: global MEMORY.md, global daily date file, project MEMORY.md, and project daily date file. Each injected frame carries a fixed marker and scans session history for deduplication, so new sessions, resumes, sub-agents, and forks never receive duplicates. The byte budget defaults to 16384 bytes and is independent of the 64KB instruction chain; on overflow the least-specific file is dropped whole first, and the most-specific file is truncated last with a notice. Sensitive delimiters inside memory content are escaped to prevent frame escape, and any internal error degrades gracefully to skipping injection without interrupting the session. The plugin is strictly read-only and never writes back.

Ideal for developers who maintain a two-level memory convention in their DSH workflows and want guaranteed context injection from the harness rather than a prompt instruction to please read the notes. If your team already has a memory directory layout, this plugin ensures every new session starts with full context out of the box, no extra prompting needed.