This project addresses the challenge of managing and leveraging the vast, unstructured session logs generated by AI coding agents. It provides a zero-dependency, local-first memory layer that makes these histories searchable and actionable, benefiting developers and researchers who rely on AI assistants for coding tasks by enabling efficient recall of past solutions and design decisions.

How It Works

Deja-vu operates as a single, statically linked Go binary that indexes session logs from supported coding agents into a local inverted index stored in ~/.cache/deja. It parses various formats (JSONL, SQLite) and crucially redacts sensitive information like API keys and JWTs at index time. This approach ensures data privacy and security while enabling fast, lexical search (7-9 ms typical) across gigabytes of historical conversations. The system also facilitates agent recall through an MCP server, automatic context injection at session start, and tools for sharing sanitized session digests or syncing memory across machines.

Quick Start & Requirements

Installation is straightforward via a shell script (curl -fsSL https://raw.githubusercontent.com/vshulcz/deja-vu/main/install.sh | sh), Go (go install github.com/vshulcz/deja-vu/cmd/deja@latest), npm (npx @vshulcz/deja-vu "query"), or Homebrew (brew install vshulcz/tap/deja-vu). It requires no external models or services, running entirely locally. sqlite3 CLI is a dependency for some agent integrations. Integration with agents is achieved via deja install --all or deja install --auto.

Highlighted Details

• Retroactive Indexing: Indexes existing agent logs, including months of history prior to installation.
• Secret Redaction: Automatically strips API keys, JWTs, and private keys at index time, replacing them with searchable placeholders.
• Fast Search: Achieves typical search latencies of 7-9 ms over gigabytes of indexed data.
• Agent Recall: Integrates with multiple coding agents (Claude Code, Codex, Cursor, etc.) for seamless recall.
• Auto-Context: Injects relevant past solutions into agent prompts automatically upon session start.
• Sharing & Syncing: Allows exporting sanitized session digests for colleagues and syncing memory stores across machines via shared folders or SSH.

Maintenance & Community

The project is primarily maintained by Vladislav Shulcz. The README does not detail specific community channels (like Discord/Slack), sponsorships, or partnerships, but provides links to installation scripts and GitHub releases for updates.

Licensing & Compatibility

The project is released under the permissive MIT license, allowing for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

While Windows builds exist, macOS and Linux are the primary battle-tested platforms. Project-specific exclusion from indexing is not yet supported, requiring manual filtering of data roots as a workaround. Subagent transcripts are excluded by default, though this behavior is configurable.