raiyanyahya/recall

What it solves

Recall addresses the "cold-start" problem for AI coding agents like Claude Code, where every new session begins without memory of previous work. It eliminates the need to manually re-explain project goals, current progress, and open threads every time a session starts, while avoiding the high token costs and privacy risks associated with sending session history to an external LLM for summarization.

How it works

Recall maintains two local Markdown files in a .recall/ directory:

• history.md: An append-only log that captures every prompt, reply, file change, and command run during a session.
• context.md: A condensed summary of the project's current state, including goals, files touched, and next steps.

Instead of using an LLM, Recall uses a classical local Python summarizer based on TF-IDF and TextRank to extract the most central sentences from the history and generate the summary. This process happens entirely on the user's machine, requiring no API keys or network calls.

Who it’s for

Developers using Claude Code or OpenCode who want a persistent, privacy-preserving, and cost-effective way to maintain project memory across multiple sessions.

Highlights

• Zero Token Cost: Summarization is performed by a local algorithm, not an LLM call, saving usage credits.
• Fully Local & Private: No data leaves the machine; transcripts and summaries are stored locally and never sent to an external API.
• Zero Friction: No complex installation or local model setup required; the summarizer is vendored and uses standard Python libraries.
• Harness Agnostic: While built for Claude Code, it can be integrated with other agents like OpenCode because it stores memory in plain Markdown.