Compound Memory

Compound Memory

Extract patterns from the current session's JSONL transcript and save as individual memory files.

• [ ] Step 1: Extract current session transcript
• [ ] Step 2: Analyze for pattern candidates
• [ ] Step 3: Draft as memory file
• [ ] Step 4: Dedup check
• [ ] Step 5: Self-evaluate
• [ ] Step 6: Present to user
• [ ] Step 7: Write to memory

Pattern crystallizer, not note-taker. You programmatically extract the current session's tool calls and results from JSONL, identify non-trivial patterns, and save them as individual memory files with frontmatter.

Your role:

• Extract current session from JSONL via jq
• Identify debugging paths, architecture seams, workaround branches
• Draft decision trees and self-evaluate against rubric
• Present and get user confirmation before writing

Not your role:

• Saving trivial fixes or language basics
• Writing without user confirmation
• Overwriting existing memory entries

When to use:

• End of session with non-trivial debugging, workarounds, architecture decisions, or integration patterns

What NOT to extract:

• Trivial fixes, one-time issues, language basics, things already in CLAUDE.md

Avoid:

• Saving prose summaries instead of decision trees
• Skipping the dedup check against existing memory files
• Lowering the quality threshold (all 5 dimensions must be >= 4/5)
• Writing to memory without user confirmation
• Analyzing from in-context memory instead of programmatic JSONL extraction

Process

Step 1: Extract Current Session Transcript

Write the jq filter to a file first (avoids zsh != history expansion and nested quoting issues), then run it:

# Write jq filter to file to avoid shell quoting issues
cat > /tmp/memory-palace-filter.jq /dev/null | head -1)

jq -r -f /tmp/memory-palace-filter.jq "$CURRENT" > /tmp/memory-palace-session.txt 2>/tmp/memory-palace-debug.log

wc -l /tmp/memory-palace-session.txt

Read the extracted transcript.

Step 2: Analyze for Pattern Candidates

Review the extracted content for:

• Error resolution paths — root cause → diagnosis → fix
• Debugging decision trees — symptom → checks → outcomes
• Architecture/integration seams — where components meet, why boundaries exist
• Workaround branches — what works, what doesn't, why

If nothing qualifies → stop, tell user "nothing worth saving in this session."

Step 3: Draft as Memory File

Format each pattern as an individual file with frontmatter:

Filename convention: {type}_{kebab-topic}.md (e.g., feedback_shell-quoting.md)

Types: user, feedback, project, reference

---
name: [Pattern Name]
description: [when to recall + what it solves — one line, specific enough to judge relevance]
type: feedback
---

├── [trigger or entry point]
│   ├── [condition A]
│   │   ├── [outcome] → [action]
│   │   └── [alternative] → [action]
│   └── [condition B]
│       └── [outcome] → [action]
├── seam: [component] ↔ [component]
│   └── decision: [why the boundary is here]
└── pitfall: [what goes wrong if you ignore this]

Step 4: Dedup Check

Glob $MEMORY_DIR/*.md and read filenames + descriptions. If a file already covers this pattern (by filename or description overlap), either skip or propose updating the existing file.

Step 5: Self-Evaluate

Score each candidate on this rubric (threshold: all >= 4/5):

| Dimension | 1 | 3 | 5 | |-----------|---|---|---| | Specificity | Abstract, no code | Has code example | Code + edge cases | | Actionability | Unclear what to do | Main steps clear | Immediately executable | | Scope Fit | Too broad/narrow | Mostly appropriate | Trigger + content aligned | | Non-redundancy | Duplicate of existing | Partial overlap | Completely unique | | Coverage | Fraction of target | Main cases covered | Main + edge + pitfalls |

Pass: all dimensions >= 4. One re-draft pass allowed; if still fails on any dimension, discard that entry.

Step 6: Present to User

Present one AskUserQuestion per pattern so the user can approve/reject each individually. Each question uses a preview showing the full file as it'll be written.

For each candidate pattern:

AskUserQuestion:
  question: "Save as {type}_{kebab-topic}.md?"
  header: "Pattern N"
  options:
    - label: "Save"
      description: "[one-line summary of the pattern]"
      preview: |
        [Full file content including frontmatter]
    - label: "Skip"
      description: "Don't save this pattern"

Only write patterns the user explicitly approved. Do NOT proceed to Step 7 for skipped patterns.

Step 7: Write to Memory

Write each approved pattern to its own file in $MEMORY_DIR/ using the filename from Step 3.

Then update $MEMORY_DIR/MEMORY.md index — one line per file with description.

If MEMORY.md doesn't exist, create it. Format:

# Memory Index

## Patterns
- `feedback_shell-quoting.md` — Shell quoting traps in zsh/awk
- `project_auth-rewrite.md` — Auth middleware rewrite context

• One line per file: filename + description
• Don't duplicate — match by filename
• Keep under 200 lines (only first 200 load into sessions)

File Ownership

| File | Access | Purpose | |------|--------|---------| | ~/.claude/projects/-$PROJECT/*.jsonl | Read only | Current session log | | $MEMORY_DIR/*.md | Read only | Dedup check | | $MEMORY_DIR/MEMORY.md | Read/Write | Index of memory files | | $MEMORY_DIR/{type}_{topic}.md | Read/Write | Individual pattern files | | /tmp/memory-palace-session.txt | Write | Extracted transcript | | /tmp/memory-palace-debug.log | Write | jq error log |

Skip Conditions

Skip if:

• No JSONL files found in project directory
• Session is trivial (no debugging, no architecture decisions, no workarounds)
• All candidate patterns already exist in memory files

Report "Nothing worth saving in this session." and exit.

Completion Criteria

• [ ] Current session extracted via jq (not from in-context memory)
• [ ] Pattern candidates identified (or "nothing worth saving" reported)
• [ ] Decision trees drafted with branches, seams, and pitfalls
• [ ] Dedup check completed against existing memory files
• [ ] All 5 rubric dimensions scored >= 4/5
• [ ] User confirmed before writing
• [ ] Pattern written to individual file in $MEMORY_DIR/
• [ ] MEMORY.md index updated with file pointer

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: qGolem
• Source: qGolem/orc
• License: MIT

Install and usage instructions live in the source repository linked above.