Recall

Your AI conversations, organized and searchable. Never lose a session again.

Cross-project session manager for Claude Code — browse, search, version control, and anti-compaction backup, without sacrificing Claude Code's security model.

See also: MemoMind — the AI-facing counterpart that gives your AI persistent memory across sessions.

Why Recall?

There are two extremes in AI coding assistants today:

OpenClaw gives you full-computer autonomy — persistent memory across days, browser automation, 700+ skills, 24/7 background agents. But it runs with elevated system permissions, no sandbox, and autonomous command execution. Cisco's AI security team found community skills performing data exfiltration without user awareness. As one developer put it: "OpenClaw is basically Claude Code without the corporate safety guardrails."

Claude Code takes the opposite approach — sandboxed execution, human approval for critical operations, project-scoped isolation. Secure by design. But this means conversations are locked to individual project directories. You can't see, search, or resume sessions from other projects. Every project is an island.

Recall bridges this gap. It adds cross-project session management, persistent organization, and long-term memory — all while keeping Claude Code's permission model and security guarantees intact. No elevated permissions, no background processes, no autonomous execution. Just a lightweight index layer on top of Claude Code's native storage.

┌─────────────────────────────────────────────────────────────┐
│              Security ◄──────────────► Capability           │
│                                                             │
│  Claude Code          Recall              OpenClaw          │
│  (project-scoped)     (cross-project)     (full-computer)   │
│  ■■■■■■■■■■          ■■■■■■■■■           ■■■■              │
│  Security             Security             Security          │
│  ■■■■                ■■■■■■■■             ■■■■■■■■■■        │
│  Memory scope         Memory scope         Memory scope      │
└─────────────────────────────────────────────────────────────┘

The Problem Recall Solves

Claude Code stores conversation history per project directory (~/.claude/projects/<path>/). This means:

• Sessions from Project A are invisible when you're working in Project B
• Built-in /resume only shows sessions from the current directory
• No way to organize conversations by topic — they're locked to whichever folder you opened
• No cross-project knowledge transfer between related conversations
• No central place to find "that conversation from last week"

This is true across VS Code, JetBrains, Cursor, Windsurf, and terminal CLI.

What Recall Does

Recall creates a centralized session index — a management layer that maps to original session files while keeping backup copies organized by user-defined categories.

~/claude-sessions/                ← Central directory (configurable)
├── _config.json                  ← Categories & settings
├── 论文/
│   ├── BSAS论文修改_meta.json    ← Metadata + mapping to original
│   └── BSAS论文修改.jsonl        ← Full session backup
├── 代码/
│   ├── WebApp重构_meta.json
│   └── WebApp重构.jsonl
└── 学习/

Key design: Recall is a read-mostly index layer, not a replacement for Claude Code's native storage. It preserves Claude Code's security model — no elevated permissions, no background processes, no autonomous execution.

Commands

Command	Description
/recall	Show interactive action menu
/recall save	Save current session (with auto-summary v2.0)
/recall search <query>	v2.0 Semantic search (embedding-based + keyword fallback)
/recall context	v2.0 Smart context injection (auto-retrieve relevant history)
/recall browse	Visual hierarchical browsing (categories → sessions → actions)
/recall list	List all saved sessions in a table
/recall list search <keyword>	Search sessions by name, summary, or tags
/recall list recent [N]	Show N most recently modified sessions
/recall list stats	Statistics overview (counts, categories, largest session)
/recall load	Load a past session as reference context in current conversation
/recall resume	Find and resume a session from any project
/recall rename	Rename a session (bidirectional sync with original project)
/recall move	Move session to a different category
/recall history	View version history, compare, or rollback a session
/recall manage	Manage categories (add, remove, view stats)
/recall reindex	v2.0 Regenerate all summaries + rebuild search index
/recall notify setup	v2.0 Configure WeChat notifications via Server酱

Core Features

Save & Organize

Save any conversation to a central location with a custom name and category. Default categories: 学习, 生活, 代码, 算法, 论文, 工作, 杂项 — fully customizable.

Auto-update: Save a session once, then /recall save again anytime for a fresh snapshot — no re-selection needed. Recall detects the existing backup by session ID and silently updates it.

Auto-save on exit (v2.0): Sessions that have been saved at least once are automatically updated when you exit Claude Code (via SessionEnd hook). No manual save needed for already-tracked sessions.

Auto-summary (v2.0): Every save generates a structured summary and tags from the session content — extracted topics, files touched, tools used. Pure rule-based, no LLM API cost.

Filesystem-first detection: Recall identifies the current session by filesystem modification time, not sessions-index.json (which can be stale). This ensures accurate detection even when Claude Code hasn't updated its own index.

Cross-Project Context Loading (/recall load)

The killer feature. Load content from a past conversation into your current session as reference context. Two modes:

• Brief: User questions + assistant text answers only
• Detailed: Includes tool usage summaries (which files were edited, what commands ran)

This enables cross-project knowledge transfer — discuss an algorithm in Project A, then load that context while working in Project B. Like OpenClaw's persistent memory, but on-demand and without background token consumption.

Interactive Browse (/recall browse)

Visual, clickable hierarchical navigation designed for hundreds of sessions:

1. Layer 1: Category overview (with session counts)
2. Layer 2: Session list within a category (paginated, 3 per page)
3. Layer 3: Session detail + executable actions (load, resume, rename, move)

Version History & Anti-Compaction (/recall history)

The central directory is a git repository. Every /recall save auto-commits a snapshot, giving you full version history:

• View history: See every save point with timestamp and message count
• Compare versions: See how a conversation grew between two saves
• Rollback: Restore any historical version — the current version is never lost
• Anti-compaction: Claude Code compresses long conversations, permanently losing early messages. Frequent /recall save captures the full conversation before compaction happens. If compaction truncates your session, rollback to a pre-compaction snapshot to recover the complete history.

Search & Statistics

• Search: Find sessions by keyword across name, summary, first prompt, and tags
• Recent: Quick view of your N most recently modified sessions
• Stats: Overview of total sessions, messages, per-category counts, most active category, largest session

Semantic Search (v2.0)

Go beyond keyword matching. /recall search uses OpenAI embeddings (text-embedding-3-small, ~$0.02/M tokens) to find sessions by meaning, not just exact words. Ask "last week's bug discussion" and get results even if you never used those exact words. Falls back to smart keyword search when no API key is configured.

Smart Context Injection (v2.0)

Say "根据之前的讨论" or use /recall context — Recall automatically retrieves the top-3 most relevant historical sessions and injects their summaries (~600 tokens total) into your current conversation. Like having a memory that remembers what you discussed across all projects.

Task Completion Notifications (v2.0)

Configure WeChat notifications via Server酱 (free). When a Claude Code task is marked complete, get a push notification on your phone. Never miss a long-running task finishing. Setup: /recall notify setup.

Resume from Anywhere

Select a session from any project, and Recall handles the rest:

• Same project: Shows the claude --resume command directly
• Different project: Opens a new VSCode window or new terminal at the target project, with the resume command ready to paste
• Original deleted: Restores from backup first, then provides the resume command

No more hunting through directories to find where a conversation lives.

Bidirectional Rename

When you rename a session in Recall, it updates both the central index and the original project's sessions-index.json. No desync between what Recall shows and what /resume shows.

Recall + MemoMind: Two Sides of the Same Coin

AI memory serves two audiences — you and your AI. Recall and MemoMind cover both:

	Recall	MemoMind
Memory for	You (the human)	The AI
Purpose	Browse, search, manage conversation history	AI remembers your preferences and context
Format	Full conversation archives (JSONL, human-readable)	Structured facts + knowledge graph + vectors
Interaction	You run /recall save, /recall search, /recall load	AI autonomously calls retain / recall / reflect
Storage	Files + SQLite index (Windows native)	PostgreSQL + pgvector (WSL)
Key value	You never lose a conversation	AI gets smarter over time

They're complementary. Recall preserves the full context so you can review what happened. MemoMind distills the essential knowledge so the AI can act on it. Use both for the complete experience.

Recall vs. OpenClaw vs. Built-in

Capability	Built-in /resume	Recall	OpenClaw
Cross-project sessions	No	Yes	Yes
Organize by category	No	Yes	Partial
Search all conversations	No	Yes (semantic v2.0)	Yes
Load past context	No	Yes	Yes (persistent)
Smart context injection	No	Yes (v2.0)	Yes
Auto-save on exit	No	Yes (v2.0 hook)	Yes
Session summaries	No	Yes (v2.0 auto)	Partial
Task notifications	No	Yes (v2.0 WeChat)	Partial
Resume from any directory	No	Yes	Yes
Central backup	No	Yes	Yes
Version history & rollback	No	Yes (git)	No
Anti-compaction	No	Yes	Partial
Sandboxed execution	Yes	Yes	No
Human approval required	Yes	Yes	No
No background processes	Yes	Yes	No
No elevated permissions	Yes	Yes	No
Works offline	Yes	Yes	Partial
Zero extra API cost	Yes	Yes*	No (24/7 tokens)

* Semantic search optionally uses OpenAI embeddings (~$0.02/M tokens, negligible cost). All other features are zero-cost.

Installation

From Skills CLI (Recommended)

npx skills add 24kchengYe/Recall

Manual Installation

# macOS / Linux
git clone https://github.com/24kchengYe/Recall.git ~/.claude/skills/recall

# Windows
git clone https://github.com/24kchengYe/Recall.git %USERPROFILE%\.claude\skills\recall

Requirements

• Claude Code (VS Code, JetBrains, Cursor, Windsurf, or terminal CLI)
• Python 3.8+ (for the helper script that parses .jsonl session files)
• Git (for version history — auto-initialized on first save)

Configuration

On first /recall save, Recall initializes:

• Storage path: Default ~/claude-sessions/ — you'll be prompted to confirm or customize
• Categories: Default set provided, fully customizable via /recall manage

All settings stored in <basePath>/_config.json.

How Claude Code Stores Sessions (Background)

For those curious about the internals:

• Each project maps to a directory: ~/.claude/projects/<encoded-path>/
• Sessions are stored as .jsonl files (one JSON object per line)
• A sessions-index.json per project tracks session metadata (but is often stale — not all sessions are indexed)
• Message types include: user, assistant, system, progress, and more
• Context compression creates compact_boundary markers in long sessions

Recall reads these native files and builds an index on top of them — it doesn't replace Claude Code's storage, just adds a cross-project management layer. When sessions-index.json is stale, Recall falls back to filesystem modification times for reliable session detection.

Changelog

v2.0.0 — 2026-03-03

• feat: Auto-save via SessionEnd hook — previously saved sessions auto-update on exit
• feat: Session summary generation — structured summaries + tags extracted from JSONL content
• feat: Semantic search with OpenAI embeddings (/recall search) — cosine similarity over session embeddings, keyword fallback
• feat: Smart context injection (/recall context) — auto-retrieve relevant history based on current conversation topic
• feat: WeChat task completion notifications via Server酱 (/recall notify setup) — TaskCompleted hook integration
• feat: Reindex command (/recall reindex) — batch regenerate summaries and rebuild search index for all sessions
• feat: SQLite search index (_index.sqlite) for fast embedding-based retrieval

v1.5.0 — 2026-03-03

• feat: Git-based version history — central directory is now a git repo, every save/rename/move auto-commits
• feat: Add /recall history command — view version timeline, compare versions, rollback to any snapshot
• feat: Anti-compaction — frequent saves preserve full conversation content before Claude Code compresses it

v1.4.0 — 2026-03-02

• fix: Use PowerShell syntax for Windows Terminal resume command (previously used CMD syntax)
• fix: Add $LOCALAPPDATA fallback when wt is not in PATH
• feat: Auto-detect terminal environment (VS Code, Windows Terminal, plain terminal) for resume
• fix: Clear CLAUDECODE env var before resume to prevent nested session errors

v1.3.0 — 2026-03-02

• docs: Rewrite README — position Recall between Claude Code and OpenClaw
• feat: Improve save reliability and cross-project resume

v1.2.0 — 2026-03-01

• feat: Add /recall browse interactive hierarchical navigation
• feat: Enhance /recall list with search, recent, and stats subcommands

v1.0.0 — 2026-03-01

• feat: Initial release
• Cross-project session save, load, resume, rename, move
• Central session index with category-based organization
• Filesystem-first session detection

License

MIT

---