Conciseness

The body is an imperative, dense protocol with no explanations of concepts Claude already knows; every section adds Engram-specific rules. Minor redundancy (the passive-capture 'safety net, not a replacement' point appears twice, and 'Do not skip step 1. Without it...' restates the preceding line) keeps it below a 5.

Actionability

Concrete tool names, a step-by-step search escalation ('First call mem_context... If not found, call mem_search... FTS5'), a full mem_save field spec with example titles, and a runnable fallback command ('engram setup claude-code'). Not 5 because a complete example mem_save call payload and several tool parameters are left unspecified.

Workflow Clarity

Each multi-step process is explicitly sequenced with checkpoints and error-recovery branches: search escalation (1→2→3), topic-update rules mandating read-before-overwrite verification ('On ambiguous identity or failed lookup, do not overwrite', 'A failed read is not evidence of absence'), failure handling in the delivery guarantee, and ordered compaction recovery with 'Only THEN continue'. Validation precedes the potentially destructive mem_update overwrite, so no cap applies.

Progressive Disclosure

A single-file skill with clear section headers, no buried or nested references, and nothing that requires external files to navigate. Not 5 because at ~101 lines it exceeds the under-50-line simple-skill exception and some content (the tool inventory, the fallback/troubleshooting details) could be offloaded to reference files.

Specificity

Names the domain ('Selective persistent memory') and 1-2 actions ('preserve verified, durable knowledge', 'safe topic updates'), but 'host-owned lifecycle' is abstract and core actions like save/search/recall are not mentioned, so coverage is not comprehensive.

Completeness

The 'what' is clear ('preserve verified, durable knowledge worth reconstructing, with safe topic updates'), but there is no 'Use when...' clause or concrete trigger situations — 'ALWAYS ACTIVE' is a degenerate trigger that names no user context. Per the rubric guideline, missing trigger guidance caps completeness at 3; not 4 because 'ALWAYS ACTIVE' does not identify when a user needs this skill.

Trigger Term Quality

'persistent memory' is the only user-natural keyword; the natural phrases users actually say ('remember', 'recall', 'across sessions', 'what did we do') are entirely absent. Not 3 because no synonyms or variations of the one keyword are present.

Distinctiveness Conflict Risk

'Selective persistent memory... safe topic updates... host-owned lifecycle' carves a distinct niche with only minor overlap risk against other memory/note-taking skills. Not 5 because the absence of distinct trigger phrases leaves some conflict risk with generic memory skills.