If you're bouncing between Cursor, Claude Code, OpenAI Codex CLI, and pi for coding assistance, you've probably lost track of useful past conversations. A new open-source tool called Recall aims to solve that by building a fast, local search index across all your AI chat histories.

What Recall Does

Recall indexes conversations from four major AI coding tools—Cursor, Claude Code, OpenAI Codex CLI, and pi—straight from their native storage locations. It doesn't move, copy, or modify any of your data. Instead, it builds a tiny SQLite database with FTS5 full-text search capabilities at ~/.recall/index.sqlite. The tool is read-only by design: your original conversation files remain the source of truth. The developer explains in the GitHub README that Recall "builds a tiny searchable index over excerpts and metadata" and lets you grep everything from your terminal. On first run, it scans all sources and builds the index—roughly one minute on real data, according to documentation showing an example with 2,597 total sessions across Cursor (1,514), Claude Code (25), and Codex (1,058).

Technical Architecture

Built in pure Go without CGO dependencies, Recall compiles down to a single static binary around 10MB. It uses SQLite FTS5 for full-text search with bm25 ranking and applies a small recency boost so recent conversations rank higher by default. The index stores excerpts trimmed to approximately 1.5KB per message along with metadata like session timestamps, project folders, and source identifiers. Storage formats vary by tool: Cursor stores data in state.vscdb files (both global and per-workspace), Claude Code uses ~/.claude/projects//*.jsonl files, Codex CLI stores sessions at ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl, and pi saves to ~/.pi/agent/sessions/. The index is disposable—nuke it with rm -rf ~/.recall/ and rebuild anytime.

MCP Integration for AI Agents

Recall includes a built-in Model Context Protocol server that exposes recall_search, recall_transcript, recall_sessions, and recall_related tools. Any MCP-capable harness can use these to let AI agents search your chat history without copy-pasting. Configuration is straightforward: Claude Code users run claude mcp add recall -- recall mcp, Codex CLI users add it to ~/.codex/config.toml, and Cursor/Cline/Windsurf users update their mcp.json file. There's also a pi-specific extension (npm:@pratikgajjar/pi-recall) that doesn't require MCP. For agents with skills systems like Claude Code Agent Skills or pi skills, installing the recall skill teaches the agent to shell out to the CLI directly—meaning you can ask things like "use recall to find how we fixed the import cycle" and watch it search autonomously.

Current Status and Installation

Recall is at v0.1.0 according to the doctor command output shown in documentation. Incremental append-only indexing, the MCP server, and the pi extension all work. Planned features include a filesystem watcher for real-time updates and a TUI for browsing results. The project is MIT licensed. Installation options include Homebrew (brew install pratikgajjar/tap/recall), a one-line shell installer, go install github.com/pratikgajjar/recall@latest, or downloading prebuilt binaries from the releases page. After installation, run recall index once to build your searchable archive, then use commands like recall for search, recall last --repo ~/code/acme-api to dump recent sessions, and recall doctor for health checks.

Key Takeaways

  • Recall creates a local SQLite FTS5 index across Cursor, Claude Code, Codex CLI, and pi chat histories without copying data
  • Pure Go implementation produces a single ~10MB static binary with no CGO dependencies
  • Built-in MCP server enables AI agents to search your conversation history autonomously
  • The index is fully disposable—rebuild anytime by deleting ~/.recall/ and running recall index

The Bottom Line

This is exactly the kind of tool that should have existed years ago. Most developers I know have abandoned useful solutions in their Cursor or Claude Code histories because there was no way to find them later. Recall fixes that without touching your data or requiring cloud sync. At v0.1 it's clearly early, but the architecture choices—read-only sources, disposable index, MCP-as-peer pattern—suggest this could become a standard part of any AI-assisted development workflow.