This project addresses the common AI problem of forgetting information between sessions and the inefficiency of Retrieval Augmented Generation (RAG) by providing a 100% local cognitive memory server for Claude. It targets users seeking persistent, intelligent, and private memory for their AI interactions, offering benefits like reduced token bloat through natural memory decay and enhanced context recall.

How It Works

Vestige implements a cognitive memory architecture inspired by 130 years of memory research. It utilizes FSRS-6 spaced repetition to naturally fade unused memories, preventing bloat. Spreading activation and synaptic tagging are employed for more nuanced memory recall. A key innovation is Prediction Error Gating, which intelligently decides whether to CREATE, UPDATE, or SUPERSEDE memories, ensuring only relevant context is retained and improving retrieval efficiency. The entire system operates locally after initial setup, prioritizing user privacy.

Quick Start & Requirements

• Primary install / run command: Download pre-compiled binaries for macOS (Apple Silicon/Intel) and Linux via curl and tar, then move executables to /usr/local/bin/. Windows users download from the Releases page. Build from source using git clone https://github.com/samvallad33/vestige && cd vestige && cargo build --release.
• Connect to Claude: Execute claude mcp add vestige vestige-mcp -s user.
• Non-default prerequisites: Internet connection required for initial model download (~130MB). cargo is needed for building from source.
• Links: Releases, FAQ, How It Works, Storage Modes, CLAUDE.md Setup, Configuration, Changelog.

Highlighted Details

• FSRS-6 spaced repetition for natural memory decay and bloat prevention.
• Spreading activation and synaptic tagging for cognitive memory modeling.
• Prediction Error Gating for intelligent memory management (CREATE/UPDATE/SUPERSEDE).
• Unified search (keyword, semantic, hybrid) and smart ingestion with duplicate detection.
• 100% local operation ensures data privacy and reduces reliance on external APIs.

Maintenance & Community

Developed by @samvallad33. Contributions are welcomed via Issues and Pull Requests, with guidance available in CONTRIBUTING.md. No explicit community channels like Discord or Slack are listed.

Licensing & Compatibility

MIT OR Apache-2.0 (dual-licensed). These permissive licenses generally allow for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

The primary integration target is Claude. An internet connection is necessary for the initial download of the embedding model; users behind proxies may need to configure HTTPS_PROXY. Troubleshooting sections address potential PATH issues and model cache location management.