Recall addresses the "cold-start" problem for users running Claude Code locally, where sessions begin without context, necessitating repetitive explanations and wasting valuable tokens. It provides a fully local, offline project memory system that logs sessions and automatically condenses them into a compact summary. This solution is ideal for users on Claude Code subscriptions seeking to preserve session context efficiently, saving both time and computational resources without incurring additional costs or compromising privacy.

How It Works

Recall operates via local hooks that capture session activity, including prompts, replies, executed commands, and touched files, appending it to a history.md log. Upon session completion or manual invocation (/recall:save), a local summarization engine employs TF-IDF and TextRank algorithms to distill the session's essence into a concise context.md file. This summary includes the project goal, key takeaways, open threads, files modified, and the last known state. The summarization is performed entirely offline using standard Python libraries, with optional Numpy acceleration, ensuring no external API calls or token expenditure for memory generation.

Quick Start & Requirements

• Installation: Install via the Claude marketplace: /plugin marketplace add raiyanyahya/recall followed by /plugin install recall@recall. Alternatively, for local development, use claude --plugin-dir /path/to/recall.
• Prerequisites: No pip install is required for core functionality. The system relies on standard Python libraries; Numpy is an optional performance enhancement. It operates entirely offline.
• Links: Official installation commands are provided. Further details on privacy are available in PRIVACY.md.

Highlighted Details

• Fully Local & Offline Operation: Guarantees that no session data or credentials leave the user's machine.
• Zero Token Cost for Memory: Session summarization is handled by local algorithms, not LLM calls, preserving subscription credits.
• Significant Token Savings: Resuming sessions from a compact context.md (~1-2K tokens) drastically reduces token usage compared to re-explaining context.
• Enhanced Privacy: Transcripts are never transmitted externally. A best-effort redaction process removes common secret patterns before writing logs.
• Zero-Friction Setup: Requires no external dependencies, API keys, or complex configuration, becoming active immediately upon plugin load.
• Deterministic Summarization: Utilizes TF-IDF and TextRank for consistent and reproducible session summaries.

Maintenance & Community

The project welcomes contributions via issues and pull requests, with guidelines provided in CONTRIBUTING.md. Security vulnerabilities should be reported privately as per SECURITY.md. Continuous integration pipelines are configured for automated testing, linting, and validation on every push and pull request. Specific community channels like Discord or Slack are not detailed in the provided documentation.

Licensing & Compatibility

The project includes a LICENSE file, but the specific open-source license (e.g., MIT, Apache 2.0) is not explicitly stated in the README. This omission requires further investigation for users considering commercial use or integration into proprietary systems. The tool is designed for Claude Code and operates locally, suggesting broad compatibility, though explicit commercial use clauses are absent.

Limitations & Caveats

Secret redaction is performed on a best-effort basis and is not a foolproof guarantee; users should review generated files before committing them. If the .recall/ directory is committed to version control for shared team memory, there is a potential risk of prompt injection if malicious content is introduced into context.md by an untrusted party with write access to the repository. The absence of a clearly stated license in the README is a significant caveat for adoption decisions.