memory-conventionlisted

# memory-convention Each consuming project grows durable, curated memory from its own review/fix/CI history. This skill defines how agents **consume** that memory: the per-skill overlay files a pattern skill reads at load time and applies on top of its baseline rules. It is purely descriptive — agents read it for context but never "execute" it, and it never modifies the baseline pattern skills shipped by this plugin. > Runtime telemetry (per-dispatch timing, tokens, tool calls) is a separate concern, written and owned entirely by the plugin's `hooks/runtime-telemetry/` scripts. The hook scripts are self-documenting; none of it is needed to read memory, so it is out of scope here. ## Where summarized memory lives ``` $MAIN_ROOT/.claude/memory/ patterns/<pattern-skill-name>.md ← per-skill improvement overlay, written by dreaming dream-log.md ← audit trail of dreaming runs (informational) ``` `$MAIN_ROOT` is the consuming project's **main working tree** root, so every slice worktree resolves to the same location: ```bash MAIN_ROOT="$(dirname "$(git rev-parse --path-format=absolute --git-common-dir)")" MEMORY_ROOT="$MAIN_ROOT/.claude/memory" ``` The directory is auto-created (by `dream-summary-memory`); it is a gitignored working dir, never committed. Overlays are produced by the `dream-summary-memory` pass — see that skill for how they are generated; the contract for *reading* them is below. ## Overlay file shape `$MAIN_ROOT/.claude/memory/p