memory-consolidate

name	memory-consolidate
description	Full memory sweep. Runs /memory-dream first, then dedupes/merges/splits durables and refreshes the index. Three approval gates. Pass --all to extend the durable sweep across every discovered repo with team-shared memory.
argument-hint	[--all]

/memory-consolidate - Full memory sweep

You are running the /memory-consolidate command — a superset of /memory-dream. Three phases, each gated by user approval before applying.

Read the routing rules at ~/.claude/memory/README.md first if not already in context.

Scope

Default scope (no flag): the user's global personal memory (~/.claude/memory/) plus the current repo's memory if cwd is inside a git repo (<cwd-repo>/.claude/memory/).

With --all: in addition, discover every other repo on the machine that has team-shared memory and include them in Phase 2 and Phase 3. Discovery:

• Roots to scan: ~/Projects/, ~/repos/, ~/Repos/, and direct subdirs of ~ (skip dotfiles already covered as part of global memory).
• For each root, search up to depth 3. A directory qualifies as a Tier-3 repo iff it contains both .git/ (or .git file in a worktree) and .claude/memory/.
• De-dup found repos by their git-common-dir so worktrees of the same repo aren't processed twice.
• Phase 1 (dream pass) is unchanged in --all mode — dailies live only in the global Tier-1 store, there's nothing repo-local to dream over.

Phase 1 — Dream pass (dailies → durable, plus archive)

Do everything /memory-dream does — see ~/.claude/commands/memory-dream.md for full behaviour. Briefly:

• Read dailies as source; read durables (global + repo) as reference.
• Plan: promote recurring/durable bullets, archive old dailies (>30d default), skip not-yet-durable.
• Wait for approval.
• Apply (promote, archive, refresh MEMORY.md daily summary).

Phase 2 — Durable sweep

Read the durable layer in scope:

• ~/.claude/memory/*.md (top-level rule files, one concept per file)
• ~/.claude/memory/tools/*.md
• ~/.claude/memory/domain/*.md
• For each in-scope repo: <repo>/.claude/memory/*.md, <repo>/.claude/memory/tools/*.md, <repo>/.claude/memory/domain/*.md

(Scope = current repo only by default; all discovered repos with --all.)

Propose, per scope:

• Dedupe overlapping entries (same fact in two files within the same scope → merge into the better-fit one). Do not dedupe across scopes — a fact in global vs. a fact in a repo can legitimately differ.
• Merge related entries within a file (consolidate sections).
• Split files that exceed ~200 lines into topic-specific siblings (e.g. tools/git.md becomes tools/git.md + tools/git-worktree.md).

Group the plan by scope (Global / <repo-a> / <repo-b> / …) so the user can approve scopes independently. Wait for approval. Apply.

Phase 3 — Index refresh

Rewrite the MEMORY.md index in each scope touched by Phase 2 to reflect current state:

• ~/.claude/memory/MEMORY.md for the global scope
• <repo>/.claude/memory/MEMORY.md for each in-scope repo

For each MEMORY.md:

• One section per topic file with a rich, keyword-dense description and a last-updated date. Sections give space for descriptions that match how the user phrases topics in conversation.
• Daily files are not indexed.
• Note last-consolidated date at the top.

If a MEMORY.md doesn't exist yet in a scope that has topic files, create it.

Present the proposed indexes grouped by scope. Wait for approval. Apply.

Constraints

• Three approval gates — do not skip any.
• Do not modify ~/.claude/projects/<mapped-cwd>/memory/ — Anthropic's auto-memory, hands off.
• Do not delete daily files even if they look stale.
• If a phase has nothing to do, report "no changes" and move to the next phase.