Wiki Memory

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: > > - Add raw/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-ingest on 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/.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/.md
 └──────────────┘
        │
        ▼
 raw/sessions/.md  ←── ready for manual /wiki-ingest

Anti-Pattern Warnings

• Do not edit settings.json manually while hooks are active — use disable then enable to 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.log and 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 into settings.json by enable-hooks.sh
• references/memory-schema.md — documents the raw/sessions/ directory layout and .state.json format

Related Skills

• /wiki-ingest — processes raw/ 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

Source & license

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

• Author: Tedydev-web
• Source: Tedydev-web/llm-wiki-skills
• License: MIT

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