By jefface
Automatic cross-session memory for Claude Code: the live model keeps a per-project memory file (decisions/paths/todos) updated, auto-loads it at session start, nudges when context fills, and backs up transcripts on session end. No MCP, no Node. Turn the whole thing off with: claude plugin disable convo-keeper.
Claude Code 套件 · v0.1.0 · 純 markdown + 3 支極小 PowerShell hook(無 Node、無 MCP)
讓 Claude Code 跨對話不失憶、又不揹著肥歷史:由「正在跟你聊、已經懂全部的當前 Claude」 把每個專案的關鍵重點寫進記憶檔,下次開場自動載入。便宜、自動、可一鍵整段關閉。
想懂原理與設計取捨(給其他人看的白話版):見 HOW-IT-WORKS.md。
| 時機(hook) | 行為 |
|---|---|
| SessionStart | 載入本專案記憶檔 → 注入給新的 Claude;同時注入「何時自動更新記憶」的規則 |
| UserPromptSubmit | 讀 transcript 估算 context 佔用,過門檻(預設 ~120k token / 60%)每個 session 提醒一次:更新記憶 + 建議 /compact |
| SessionEnd | 把原始 transcript 複製到備份資料夾(保險,免費,不做摘要) |
記憶由當前 Claude 寫(被提醒時,或你打 /keep)——不另開 claude -p、不需 claude setup-token、不多花錢。
claude plugin marketplace add C:\jeffWorkspace\project\jefface\AI\convo-keeper
claude plugin install convo-keeper@convo-keeper-local
# 之後完全重啟 Claude Code,hook 才會載入
claude plugin disable convo-keeper # 關閉全部(下個 session 起生效)
claude plugin enable convo-keeper # 重新開啟
disable 後,目前這條對話會跑到結束,下一條才完全靜默(hook 在 session 開場才註冊)。
| 路徑 | 用途 |
|---|---|
~/.claude/convo-keeper/<專案key>.md | 每專案記憶檔(決策/路徑/待辦) |
~/.claude/convo-keeper/transcripts/ | SessionEnd 的原始 transcript 備份 |
~/.claude/convo-keeper/.nudged-<sid> | 每 session 只提醒一次的旗標 |
手動存檔:要我立刻把「決策/路徑/待辦」寫進本專案記憶檔。平常不打也行—— context 到 60% 時我會被提醒自動更新。
/compact。官方規格:沒有任何 hook 輸出能執行 slash 指令。
所以本套件是「偵測 + 提醒」,最後那一下 /compact 仍要你按。usage(input + cache_read + cache_creation)
除以 200k 估算;若你的模型視窗不是 200k,百分比會偏移,門檻請自行調 ctx-nudge.ps1 的 $THRESHOLD。$ErrorActionPreference=SilentlyContinue,出錯一律安靜跳過,不擋你工作。把這個資料夾 push 到 git repo,另一台:
claude plugin marketplace add <你的 git url>
claude plugin install convo-keeper@convo-keeper-local
hook 內路徑全用 ${CLAUDE_PLUGIN_ROOT} 與 $env:USERPROFILE,不寫死,可移植(限 Windows)。
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
npx claudepluginhub jefface/convo-keeper --plugin convo-keeperConsult multiple AI coding agents (Gemini, OpenAI, Grok, Perplexity, plus codex and gemini CLIs when installed) to get diverse perspectives on coding problems
20 SEO/GEO skills and 5 commands on one shared contract for keyword research, content creation, technical audits, schema markup, monitoring, quality gates, entity truth, and campaign memory.
Complete AI coding workflow system. Self-correcting memory + persistent FTS5-indexed research wikis + auto-research loop + multi-LLM council on a single SQLite store. 33 skills, 8 agents, 22 commands, 37 hook scripts across 24 events. Cross-agent via SkillKit.
Production-grade academic research pipeline for Claude Code: research → write → review → revise → finalize. 4 skills, 27 modes, 39-agent ensemble, v3.7.3 + v3.8 L3 claim-faithfulness gate, v3.9.0 cross-index triangulation, v3.10 triangulation policy layer, v3.11 deterministic citation verification gate (#182).
Claude + Obsidian knowledge companion. Sets up a persistent, compounding wiki vault (Karpathy's LLM Wiki pattern). v1.7 "Compound Vault" + v1.8 methodology modes close 5 of 5 priority gaps from the May 2026 compass artifact. Ships: substrate alignment with kepano/obsidian-skills, default Obsidian CLI transport, hybrid retrieval (contextual prefix + BM25 + cosine rerank per Anthropic's Sept 2024 research), per-file advisory locking for multi-writer safety, pre-commit verifier agent, AND methodology modes (LYT / PARA / Zettelkasten / Generic) for first-class organizational support no other Claude+Obsidian competitor offers. v1.7.x audit closure: every BLOCKER + HIGH + MEDIUM + LOW finding from the v1.7.0 audit is CLOSED or DEFERRED-with-rationale. Optional DragonScale Memory extension (log folds, deterministic addresses, semantic tiling lint, boundary-first autoresearch).