Add fable-mem: cross-project persistent memory system

[blyatiful1]

Introduces a machine-wide, searchable memory corpus that layers on top of Claude's native per-repo memory, closing the gap where decisions banked in one project become invisible in another.

Summary

This adds fable-mem, a new component kind consisting of:

• A CLI tool (claude/cli/mem.py) for indexing and searching memories across all projects
• Three protocol hooks for automatic memory management and privacy enforcement
• Two workflows for memory corpus hygiene and session mining
• A skill documenting the memory-search capability

The system maintains a disposable sqlite3 FTS5 index over both the global corpus (~/.claude/memory/*.md) and every native per-repo corpus (~/.claude/projects/*/memory/*.md), with graceful degradation to LIKE-based search when FTS5 is unavailable. It uses stdlib only (no pip/venv) so it runs in bare python3 environments.

Key Changes

CLI (claude/cli/mem.py)

• index: incremental mtime-based indexing with per-file resilience; --rebuild flag for full reconstruction
• search: keyword recall with scope filtering (global/project), FTS5 with LIKE fallback
• show: display memory body by ID
• stats: corpus statistics by scope and project
• doctor: health check including FTS mode detection and privacy scan
• gc-scan: mechanical detection of near-duplicates, stale entries, relative dates, and same-topic pairs

Hooks

• userpromptsubmit-mem-recall.py: injects up to 3 memory pointers (title + description + path, never bodies) as labelled reference data on every prompt; per-session deduplication; fail-open guarantee
• sessionend-mem-journal.py: appends NDJSON breadcrumb to journal.ndjson (session metadata, git state) and runs incremental reindex; bounded git subprocess budgets to stay under SessionEnd timeout
• pretool-mem-privacy-guard.py: blocks Write/Edit/MultiEdit into global corpus if pending content matches privacy.toml patterns; fail-open on any ambiguity

Workflows

• /memory-review: mines session journal for high-activity sessions that banked zero memories; proposes capture candidates via three-way judges
• /memory-gc: corpus-health sweep combining mechanical gc-scan, contradiction judges for same-topic pairs, relative-date absolutization, and index rebuild; proposals only, never deletes

Configuration

• privacy.toml: user-owned pattern file (seeded but never overwritten) for work-marker detection
• Settings snippets wired to PreToolUse (privacy guard) and UserPromptSubmit (recall hook)

Implementation Details

• BASE resolution: respects CLAUDE_DIR environment variable, falls back to ~/.claude; enables scratch-dir testing without touching $HOME
• FTS5 probing: availability checked at DB-open with graceful degradation; FABLE_MEM_FORCE_DEGRADED=1 escape hatch for testing
• WAL mode: enabled for concurrent read/write without blocking
• Corpus discovery: stable sorted iteration over global + all project memories
• Frontmatter parsing: minimal one-level YAML with nested metadata: block support; postmortem convention fields (name, description, type, created, verified, visibility)
• Relative-date detection: regex pattern for hygiene warnings (e.g., "yesterday", "last week")
• Fail-open posture: read paths (search/show/doctor) catch CorruptIndex gracefully; mutating paths (index/stats/gc-scan) recreate corrupt index; hooks exit 0 on any failure except privacy guard which exits 2 on positive match
• Deduplication: per-session state file tracks injected memory IDs to avoid repeating pointers within a session

Testing

Comprehensive test suites for each component:

• test_mem_cli.py: CLI functionality (index, search, show, stats, doctor, gc-scan)
• test_mem_recall_hook.py: memory injection, deduplication, budget enforcement
• `test_mem_privacy

https://claude.ai/code/session_01Vko5e15eEpKpFNZBJBYoDA