Distill session transcripts into durable memory. Use when processing staged sessions (after the Stop hook), when the user runs /engraim:curate or /engraim:backfill, or whenever you should turn "what just happened" into lasting knowledge. Turns raw Claude Code session JSONL into wiki pages + temporal facts, with provenance and no secrets.
Instalação
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Distill session transcripts into durable memory. Use when processing staged sessions (after the Stop hook), when the user runs /engraim:curate or /engraim:backfill, or whenever you should turn "what just happened" into lasting knowledge. Turns raw Claude Code session JSONL into wiki pages + temporal facts, with provenance and no secrets.
Curate: transcript -> memory
Goal: turn raw session transcripts into a denser, more useful knowledge base — without hoarding the transcript itself.
Inputs
Staged sessions: .engraim/pending/sessions.jsonl (each line: {session_id, transcript_path, ended_at, reason}), or a specific transcript_path.
Transcripts live under ~/.claude/projects/<url-encoded-project-path>/<session-id>.jsonl — append-only JSONL of user/assistant messages, tool calls + results, and thinking blocks.
Process (per session)
ingest_session(transcript_path) — mechanical first pass: registers the session as an immutable source and returns a digest (prompts, tool commands run, files touched, errors seen). This is scaffolding, not the distillation.
You distill from the transcript + digest:
Decisions made and why (the reasoning, not just the outcome).
Fixes that worked (problem -> root cause -> solution) — these are prime wiki-page material.
Environment facts learned (inventory, versions, conventions, quirks specific to THIS project/lab).
Skip: routine edits, dead ends with no lesson, anything already in memory.
Write it down:
wiki_upsert(page, content, links) for narrative/how-to knowledge; one page per entity/concept; cross-link with [[wikilinks]]. Fuzzy-check for an existing page before creating a near-duplicate.
remember(observation, ...) for atomic facts; set source to the session_id.
Clear the processed entries from the pending list; report a 2-3 line summary.
Rules
Distill, don't dump. A good session yields a handful of facts and 1-3 wiki edits, not a transcript copy.
No secrets. Redact/skip API keys, tokens, passwords, private URLs seen in tool output.
Provenance + scope. Cite session_id. Only ever read THIS project's transcript folder (derived from cwd), never other projects'.
No silent drops. If you decide something notable isn't worth keeping, that's fine — but don't silently lose a real decision; when in doubt, record a terse fact.
