backfill

Backfill — Populate Memory from Conversation History

Scan Claude Code conversation transcripts for a time window, extract durable observations through the consolidation engine, and seed them into the memory store.

Invocation

When this skill triggers, prompt the user for two inputs before running anything:

1. Time window (required): "How far back should I scan? (e.g. 30d, 2w, 3m)"
2. Focus area (optional): "Is there a specific topic to focus on? (e.g. 'testing patterns', 'how you prefer to communicate', 'architecture decisions') — or leave blank for general extraction. Note: focus injection is experimental and hasn't been formally evaluated against the base prompt yet."

Only proceed after the user responds. Do not assume defaults.

Usage (after gathering inputs)

/memesis:backfill 30d
/memesis:backfill 2w --focus "testing patterns and preferences"
/memesis:backfill 7d --project app --limit 20

Pipeline

Three scripts run in sequence. Each produces output consumed by the next.

Step 1: Scan

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/scan.py <duration> [--project NAME] [--limit N] [--min-size KB]

Reads JSONL transcripts from ~/.claude/projects/*/, extracts user/assistant text (stripping tool output, system messages, large code blocks), and writes session summaries to backfill-output/summaries.jsonl.

Duration formats: 30d (days), 2w (weeks), 4h (hours), 6m (months).

Step 2: Consolidate

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/consolidate.py [--limit N] [--focus "TOPIC"]

Feeds each summary through the consolidation prompt via the Anthropic API. Writes keep/prune decisions to backfill-output/consolidation-results.jsonl.

The --focus flag injects guidance into the prompt to bias extraction toward a specific topic without overriding the keep/prune criteria. Examples:

• --focus "how Emma communicates and makes decisions"
• --focus "testing patterns, CI/CD workflows, and test isolation"
• --focus "architecture decisions and their constraints"
• --focus "corrections and mistakes — what went wrong and why"

Step 3: Seed

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/seed.py [--project-context PATH] [--dry-run] [--report]

Reads kept observations from consolidation results and creates consolidated memories in the store. Deduplicates by content hash — safe to run multiple times.

Use --report to see a quality breakdown before seeding. Use --dry-run to preview without writing.

Procedure

When the user invokes /memesis:backfill <duration>, run the three scripts in sequence:

1. Parse duration and any flags from the user's input
2. Run scan.py with the duration and flags
3. Run consolidate.py (pass --focus if the user specified a topic)
4. Run seed.py --report first to show results, then seed.py to commit

If the user only wants to see what would be extracted, run steps 1-2 and then seed.py --dry-run.

Examples

/memesis:backfill 30d
/memesis:backfill 2w --focus "how this user prefers to work with AI agents"
/memesis:backfill 7d --project app