notque/vexjoy-agent / skills/meta/auto-dream
auto-dream
UnverifiedBackground memory consolidation — overnight review, merge, and injection payload for memory files.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 5.6 KB |
| dream-prompt.md | markdown | 12.2 KB |
| references/concurrency.md | markdown | 7.4 KB |
| references/dream-cycle-testing.md | markdown | 9.3 KB |
| references/headless-cron-patterns.md | markdown | 10.8 KB |
| references/logging-patterns.md | markdown | 11.5 KB |
| references/memory-file-operations.md | markdown | 12.8 KB |
references/memory-file-operations.md
View on GitHub# Memory File Operations Reference
> **Scope**: Patterns for reading, writing, merging, archiving, and indexing memory files in the toolkit's persistent memory system. Covers YAML frontmatter structure, atomic writes, MEMORY.md index management, and staleness/duplicate detection criteria. Does NOT cover the SQLite learning database.
> **Version range**: All toolkit versions using the `~/.claude/projects/.../memory/` layout
> **Generated**: 2026-04-15 — verify frontmatter field names against current memory files
---
## Overview
Memory files are the toolkit's persistent cross-session knowledge store. Each file carries a YAML frontmatter header and markdown body. The `MEMORY.md` file is a mandatory index — every memory file must have a pointer there. The auto-dream cycle reads, consolidates, and archives these files under strict rules: never delete (only archive), write MEMORY.md atomically, and flag conflicts for human review rather than auto-resolving them.
---
## Pattern Table
| Operation | Pattern | Notes |
|-----------|---------|-------|
| Create memory file | Write frontmatter + body to `memory/{type}_{slug}.md` | Add pointer to MEMORY.md immediately |
| Update MEMORY.md | Write to `.tmp`, then rename | Never write MEMORY.md directly — partial writes corrupt index |
| Archive stale file | `mv memory/file.md memory/archive/file.md` | Never `rm` — archiving is reversible, deletion is not |
| Merge duplicates | Combine into new file, archive both sources | Add `merged_from` field to merged file's frontmatter |
| Detect staleness | Check ALL THREE signals (see Stale Detection) | One or two signals is insufficient — all three required |
---
## YAML Frontmatter Structure
Every memory file must start with a YAML block:
```markdown
---
name: descriptive-memory-name
description: One-line description — used to decide relevance at session start, be specific
type: user | feedback | project | reference | insight
---
```
**Required fields**: `name`, `description`, `type`.
**Optional fields used by consolidation**:
```yaml
---
name: merged-absolute-paths
description: Use absolute paths for bash scripts and cron commands
type: feedback
merged_from:
- feedback_absolute_paths_a.md
- feedback_absolute_paths_b.md
---
```
**Why**: The `description` field is what gets injected into session context via MEMORY.md. Vague descriptions ("some notes about bash") produce useless injection payloads. The `merged_from` list preserves audit trail when consolidating duplicates.
---
## Correct Patterns
### Atomic MEMORY.md update
Never write `MEMORY.md` directly. Always write to a `.tmp` file, then rename:
```bash
# Write new index content to temp file
python3 - <<'EOF'
content = build_new_memory_index() # your index generation logic
with open("memory/MEMORY.md.tmp", "w") as f:
f.write(content)
EOF
# Atomic rename — this is the only filesystem operation that can fail atomically
mv memory/MEMORY.md.tmp memory/MEMORY.md
```
**Why**: If the process is interrupted while writing `MEMORY.md` directly, the index is partially written and unreadable. The rename (`mv`) is atomic on POSIX filesystems — the index either fully updates or stays unchanged.
---
### Archive-not-delete for stale files
```bash
# Create archive directory if needed
mkdir -p memory/archive/
# Move (never rm) stale memory to archive
mv memory/stale-project-memory.md memory/archive/stale-project-memory.md
# Remove from MEMORY.md (via atomic tmp pattern above)
```
**Why**: Deleted memories are unrecoverable if the staleness assessment was wrong. Archived files in `memory/archive/` are out of the active rotation but available for manual review. Archive is reversible; deletion is not.
---
### Merge with provenance
When two memories cover substantially the same ground, merge into a single file:
```markdown
---
name: absolute-path-preference
description: Always use absolute paths in bash scripts, cron commands, and file operations
type: feedback
merged_from:
- feedback_bash_absolute_paths.md
- feedback_cron_path_issues.md
---
Use absolute paths everywhere in bash and cron scripts. Relative paths break when
the working directory changes (e.g., cron runs from `/`, not the project root).
**Why:** Two separate incidents: one from a cron script silently doing nothing because
it couldn't find the config file, one from a hook script failing because `./scripts/`
resolved to the wrong location.
```
After writing the merged file, archive both sources and update MEMORY.md.
---
## Stale Detection Criteria
A memory is stale only when ALL THREE signals are present. One or two is insufficient.
| Signal | How to check | What counts |
|--------|-------------|-------------|
| No recent git activity | `git log --oneline -20 \| grep -i {topic}` | Topic appears in 0 of last 20 commits |
| No recent session refs | SQLite: sessions last 7 days mention topic | Topic in 0 session summaries in last 7 days |
| File age > 30 days | `stat memory/{file}.md` or file mtime | Last modified > 30 days ago |
**Example**: A project memory about "mobile release freeze ends 2026-01-15" with no matching git commits, no matching recent sessions, and file age 45 days → all three signals → candidate for archiving.
**Counter-example**: A feedback memory about "never use --no-verify" with no recent git commits about it → only ONE signal (no recent refs). Keep this guidance active until at least two corroborating signals appear.
---
## Duplicate Detection
Two memories are duplicates when a reader could replace one with the other without losing information. Match on semantic content similarity, not filenames.
```bash
# Detection: look for memory files with overlapping descriptions in MEMORY.md
grep -n '—' memory/MEMORY.md | sort -k3 # sort by description to spot semantic overlap
```
**Examples of duplicates**:
- `feedback_use_absolute_paths.md`: "Always use absolute paths in bash"
- `feedback_bash_paths.md`: "Use absolute paths in shell scripts, not relative"
**Examples that are NOT duplicates** (different scope):
- `feedback_no_verify_hooks.md`: "Never skip pre-commit hooks"
- `feedback_test_before_commit.md`: "Run tests before committing"
---
## Pattern Catalog
<!-- no-pair-required: section heading only, paired Do instead blocks appear in each sub-entry below -->
### Write MEMORY.md via Atomic Rename
**Detection**:
```bash
# Find code that writes MEMORY.md without going through a .tmp file
grep -rn 'MEMORY\.md' hooks/ scripts/ | grep -v '\.tmp' | grep -v 'read\|cat\|grep\|open.*r'
rg 'write.*MEMORY\.md(?!\.tmp)' --type py
```
**Signal**:
<!-- no-pair-required: this is the detection sub-block inside a code fence; Do instead appears in the enclosing failure mode entry -->
```python
with open("memory/MEMORY.md", "w") as f:
f.write(new_index_content)
# If interrupted here, MEMORY.md is partially written
```
**Why this matters**: A partial write leaves MEMORY.md in an invalid state. The session-start hook reads this file to load context — a corrupted index means no memory injection at all until manually repaired.
**Preferred action**: Write all MEMORY.md updates to a `.tmp` file first, then use `os.rename()` to atomically replace the live index. The rename either fully succeeds or leaves the original intact — no partial state is possible.
**Preferred action**:
```python
with open("memory/MEMORY.md.tmp", "w") as f:
f.write(new_index_content)
os.rename("memory/MEMORY.md.tmp", "memory/MEMORY.md")
```
---
### Archive Memory Files Instead of Deleting
**Detection**:
```bash
# Find any rm operations on memory files
grep -rn 'rm.*memory/' scripts/ hooks/
rg 'os\.remove|os\.unlink|Path.*unlink' --type py | grep -i 'memory'
```
**Signal**:
<!-- no-pair-required: this is the detection sub-block inside a code fence; Do instead appears in the enclosing failure mode entry -->
```bash
rm memory/stale-project-memory.md # Unrecoverable if assessment was wrong
```
**Why this matters**: Staleness assessment can be wrong. A memory that looks inactive (no recent git refs, no session refs) might still be valid ongoing guidance. Archive preserves reversibility; deletion does not. Dream cycle design rule: "Never delete files."
**Preferred action**: Move stale files to `memory/archive/` rather than deleting them. Archived files are out of active rotation but available for manual review if the staleness assessment turns out to be wrong.
**Preferred action**:
```bash
mkdir -p memory/archive/
mv memory/stale-project-memory.md memory/archive/stale-project-memory.md
```
---
### Flag Contradictory Memories for Human Review
**Detection**:
```bash
# No automated detection — this is a process violation, not a code pattern
# Review: does the consolidation code ever rewrite one memory based on another?
grep -rn 'conflict\|contradict' scripts/ | grep -v 'flag\|report\|human'
```
**Signal**:
<!-- no-pair-required: this is the detection sub-block inside a code fence; Do instead appears in the enclosing failure mode entry -->
```
Memory A: "Use ruff for Python linting"
Memory B: "Use flake8 for Python linting"
→ Auto-consolidation picks A because it's newer and deletes B
```
**Why this matters**: The conflict may reflect a real, unresolved project decision. Auto-resolution buries the disagreement. The human reviewer needs to see both to make the call.
**Preferred action**: When CONSOLIDATE finds contradictory memories, leave both files untouched and record the conflict in the dream report under "Conflicts Requiring Review" with the exact contradiction described. Human judgment resolves the conflict on the next review cycle.
**Preferred action**: Leave both files untouched. Record in the dream report under "Conflicts Requiring Review" with the specific contradiction described. Do not modify either file.
---
### Keep MEMORY.md Index Under 200 Lines
**Detection**:
```bash
wc -l memory/MEMORY.md
# Should be under 200 lines; every line after 200 is truncated in context injection
```
**Signal**:
<!-- no-pair-required: this is the detection sub-block inside a code fence; Do instead appears in the enclosing failure mode entry -->
An index that grows unbounded as new memories are added without old ones being consolidated or archived, eventually exceeding the 200-line truncation limit.
**Why this matters**: The session-start hook injects MEMORY.md contents. Context injection is truncated at 200 lines — memories after that line are silently excluded from session context, defeating the purpose of writing them.
**Preferred action**: Treat MEMORY.md line count as a managed resource. Each dream cycle should consolidate or archive at least as many memories as it adds. When the index approaches 150 lines, prioritize archiving stale project memories before adding new ones.
**Preferred action**: Each dream cycle should consolidate at least as many memories as it adds. If MEMORY.md approaches 150 lines, prioritize archiving stale project memories in the next cycle.
---
## Error-Fix Mappings
| Error / Symptom | Root Cause | Fix |
|-----------------|------------|-----|
| Session starts with no memory context | MEMORY.md corrupted (partial write) | Restore from `memory/archive/` or `MEMORY.md.tmp` if it exists |
| Memory file missing frontmatter | Written without YAML block | Add `---\nname: ...\ndescription: ...\ntype: ...\n---\n` header |
| Merged memory loses original context | `merged_from` field not set | Add `merged_from: [source_a.md, source_b.md]` to merged frontmatter |
| Memory pointer in MEMORY.md has no corresponding file | File was deleted rather than archived | Check `memory/archive/` for the file; restore if needed |
| Stale memory archived that should have stayed | Only 1-2 staleness signals, not all 3 | Restore from archive; require all 3 signals before archiving |
| MEMORY.md index entries missing | MEMORY.md.tmp rename failed | Check for `.tmp` file and rename manually; re-run with corrected atomic write |
---
## Detection Commands Reference
```bash
# Check MEMORY.md line count (should be under 200)
wc -l memory/MEMORY.md
# Find non-atomic MEMORY.md writes
grep -rn 'MEMORY\.md' scripts/ hooks/ | grep -v '\.tmp\|read\|cat\|grep'
# Find rm operations on memory files
grep -rn 'rm.*memory/' scripts/ hooks/
# Verify all MEMORY.md pointers have corresponding files
python3 -c "
import re, os
content = open('memory/MEMORY.md').read()
links = re.findall(r'\[.*?\]\((.*?\.md)\)', content)
missing = [l for l in links if not os.path.exists(f'memory/{l}')]
print('Missing:', missing or 'none')
"
# Find memory files without YAML frontmatter
grep -rL '^---' memory/*.md 2>/dev/null | grep -v 'MEMORY\|archive'
# Count memories by type
grep -h '^type:' memory/*.md 2>/dev/null | sort | uniq -c
```
---
## See Also
- `skills/meta/auto-dream/references/headless-cron-patterns.md` — wrapper script and cron invocation patterns
- `skills/meta/auto-dream/dream-prompt.md` — Phase 3 CONSOLIDATE and Phase 4 SYNTHESIZE instructions
- Memory system design: `skills/meta/auto-dream/SKILL.md` — Safety constraints section
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |