<< All versions
Skill v1.0.0
currentLLM-judged scan90/100tedydev-web/llm-wiki-skills/wiki-memory
──Details
PublishedOctober 3, 2026 at 02:43 PM
Content Hashsha256:95d470e59cc1cd63...
Git SHA
──Files
Files (1 file, 4.3 KB)
SKILL.md4.3 KBactive
SKILL.md · 90 lines · 4.3 KB
version: "1.0.0" name: wiki-memory description: "Optional add-on: capture Claude Code session transcripts as wiki sources for later ingest. Off by default. Subcommands: enable, disable, status, flush, logs." allowed-tools: Bash Read Write Glob Grep
wiki-memory — Session Transcript Capture
Optional add-on skill that registers Claude Code lifecycle hooks to capture session transcripts into your LLM Wiki vault's raw source directory. Captured transcripts become ingest candidates for /wiki-ingest. The skill is off by default — nothing runs until you explicitly enable it.
Privacy Notice
WARNING: Captured transcripts contain your full conversation with the LLM, including any secrets, API keys, passwords, or personal data you have pasted into the session. Before enabling this skill:- Addraw/sessions/to your vault's.gitignore(and.obsidianignore) before enabling capture.- Never commit transcript files to a public or shared repository.- Review captured files before running/wiki-ingeston them — scrub any sensitive content first.- If you work in a shared environment, confirm that your vault storage path is not synced to a shared drive while capture is active.You accept full responsibility for the content of any captured transcripts.
Activation
/wiki-memory enable [--scope global|project]
--scope global— registers hooks in~/.claude/settings.json(affects all projects)--scope project— registers hooks in./.claude/settings.json(current project only, default)
Auto-compile on SessionEnd is deferred to v1.2 — for now, run /wiki-ingest manually after reviewing transcripts.
Subcommands
| Subcommand | Script | Description | |
|---|---|---|---|
enable | scripts/enable-hooks.sh | Register 3 lifecycle hooks in the target settings.json | |
disable | scripts/disable-hooks.sh | Remove hooks from settings.json; leave captured files intact | |
status | scripts/status.sh | Show whether hooks are registered + last capture timestamp | |
flush | scripts/flush-session.sh | Force-write the current in-progress session buffer to disk | |
logs | scripts/logs.sh | Tail or page the capture log for the current project/global scope |
Capture Flow
Claude Code lifecycle│▼┌──────────────┐ SessionStart hook│ hook-session │──► create session buffer file at raw/sessions/<session-id>.md│ -start.sh │└──────────────┘│(conversation in progress)│▼┌──────────────┐ PreCompact hook (non-blocking)│ hook-pre- │──► snapshot current transcript to buffer│ compact.sh │ (must not throw — isolated error handling)└──────────────┘│▼┌──────────────┐ SessionEnd hook│ hook-session │──► finalize transcript file, append metadata frontmatter,│ -end.sh │ write to raw/sessions/<session-id>.md└──────────────┘│▼raw/sessions/<session-id>.md ←── ready for manual /wiki-ingest
Anti-Pattern Warnings
- Do not edit `settings.json` manually while hooks are active — use
disablethenenableto regenerate entries cleanly. - PreCompact hook must never block — the hook implementation uses isolated error trapping; a failure in snapshot logic will log to
.wiki-memory.logand exit 0 so Claude Code's compaction proceeds unaffected. - Do not run `flush` mid-sentence — flush captures the buffer as-is; an incomplete exchange will appear truncated in the transcript.
- Scope mismatch — if you enabled globally but want per-project overrides, disable globally first to avoid duplicate hook entries.
Related Files
references/hooks-template.json— hook entry template merged intosettings.jsonbyenable-hooks.shreferences/memory-schema.md— documents theraw/sessions/directory layout and.state.jsonformat
Related Skills
/wiki-ingest— processesraw/files (including captured transcripts) into structured wiki pages/wiki— initial vault setup (run before enabling wiki-memory)/wiki-query— query and synthesize from compiled wiki pages