- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add Ar9av/obsidian-wiki --skill memory-bridge -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add Ar9av/obsidian-wiki --skill memory-bridge -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add Ar9av/obsidian-wiki --skill memory-bridge -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add Ar9av/obsidian-wiki --skill memory-bridge -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/Ar9av/obsidian-wiki/tree/HEAD/skills/memory-bridge來自發布者
You are helping the user browse and compare their Obsidian wiki knowledge filtered by which AI tool originally produced it. The wiki tracks source provenance in .manifest.json and page sources: frontmatter — this skill surfaces that metadata as a navigable view.
llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → global config → prompt setup). This gives OBSIDIAN_VAULT_PATH.$OBSIDIAN_VAULT_PATH/.manifest.json — this is the source-of-truth for what tool produced what.$OBSIDIAN_VAULT_PATH/index.md for page titles and one-line descriptions.Parse the user's invocation to determine mode:
Recognized tool names: claude, codex, hermes, openclaw, copilot, pi, manual (hand-written), ingest (wiki-ingest documents).
Read .manifest.json. For each source entry, extract:
source_type — maps to tool name:
claude_conversation, claude_memory, claude_audit_log, claude_desktop_session → claudecodex_rollout, codex_index, codex_history → codexhermes_memory, hermes_session → hermesopenclaw_memory, openclaw_daily_note, openclaw_session, openclaw_dreams → openclawcopilot_session, copilot_checkpoint, copilot_transcript, copilot_memory_artifact → copilotpi_session → pidocument → ingestmanualpages_created and pages_updated — the wiki pages that came out of this sourceBuild a map:
tool_pages = {
"claude": set(pages created/updated by claude sources),
"codex": set(pages created/updated by codex sources),
...
}
A page can appear in multiple tools' sets if multiple tools contributed to it.
Filter tool_pages[<tool>] and present as a grouped list:
## Knowledge from <tool> (<N> pages)
### By category
- concepts/ — N pages
- entities/ — N pages
- skills/ — N pages
...
### Pages
| Page | Category | Tags | Last updated |
|------|----------|------|--------------|
| [[page-name]] | concept | tag1, tag2 | 2026-04-10 |
...
Read frontmatter for the listed pages (grep for ^(title|category|tags|updated):) — do not read full page bodies unless the user asks.
Within the filtered page set, run:
rg -l "<topic>" <pages in tool set>
Then grep section headers (^##) around matches to give context without full reads. Present results as a ranked list with the matching excerpt.
Compute:
only_in_a = tool_pages[a] − tool_pages[b]only_in_b = tool_pages[b] − tool_pages[a]shared = tool_pages[a] ∩ tool_pages[b]If no specific tools are given, compare all tools pairwise (limit to pairs with >0 overlap or unique pages to keep output concise).
Present:
## Memory Bridge Diff — <tool-a> vs <tool-b>
### Only in <tool-a> (<N> pages)
These concepts exist in your wiki from <tool-a> sessions but <tool-b> has never touched them.
<list with one-line descriptions from index.md>
### Only in <tool-b> (<N> pages)
<list>
### Shared (<N> pages)
Both tools have contributed to these pages.
<list — only show if ≤15; otherwise just the count>
### Notable gaps
<highlight the most interesting asymmetries — e.g. "codex has 12 pages on build tooling that claude has never seen">
Build a matrix showing every page and which tools have touched it. Cap at 50 rows; sort by number of contributing tools descending (most cross-tool pages first — these are the richest nodes).
| Page | claude | codex | hermes | copilot | pi |
|------|--------|-------|--------|---------|----|
| [[react-patterns]] | ✓ | ✓ | — | ✓ | — |
| [[rust-ownership]] | — | ✓ | — | — | ✓ |
After generating output, if the impl-validator skill is available in the current environment, spawn it as a subagent:
impl-validator check:
goal: "Browse/diff wiki knowledge by source tool and surface cross-tool blind spots"
artifacts: [the output you just generated]
checks:
- Did you correctly parse source_type from .manifest.json?
- Are page counts plausible (not 0 unless vault is empty)?
- Is the diff symmetric (a−b and b−a are disjoint)?
- Did you avoid reading full page bodies when not needed?
Apply any issues it surfaces before presenting output to the user.
Append to $OBSIDIAN_VAULT_PATH/log.md:
- [TIMESTAMP] MEMORY-BRIDGE mode=<browse|search|diff|map> tool=<tool> pages_shown=N
[[wikilinks]] for page references (or standard Markdown links if OBSIDIAN_LINK_FORMAT=markdown is set)..manifest.json is empty or missing, say so clearly and suggest running /wiki-history-ingest first.>
---name: memory-bridgedescription: >---# skills/memory-bridge/SKILL.md常見問題
從發布者下載 skills/memory-bridge/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/Ar9av/obsidian-wiki。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。