engrim

engrim is a groundbreaking local-first, project-scoped SQLite memory engine designed for seamless integration with various AI models, including Google Antigravity, Claude Code, Cursor MCP, and Windsurf. This repository offers a universal approach to managing episodic memory across multiple environments, enabling developers to maintain continuity in their projects without being locked into any single vendor's ecosystem.

Core Value Proposition

In an era where context windows are expanding to over a million tokens, developers contend with attention dilution—a scenario where reasoning suffers and costs increase with every interaction. engrim mitigates this issue by providing 4,000 characters of curated episodic working memory, ensuring that the architectural decisions and user constraints remain intact across model transitions.

• Vendor-Neutral Memory Management: Transition between different AI models effortlessly without losing track of decisions or project context. Developers can shift from one framework to another, ensuring that agents pick up seamlessly where their predecessors left off.
• Effortless Context Reload: The engine allows users to clear session contexts and reload necessary memories using a simple command, facilitating smooth transitions and project updates.
• Optimized Context Retrieval: By employing SQLite FTS5 for keyword searches combined with static vector embeddings, engrim delivers a fast and efficient memory retrieval system.

Proven Empirical Results

In extensive testing across 105 sessions on a complex algorithmic trading system, engrim demonstrated its reliability:

• A staggering reduction in context retrieval costs was achieved, consolidating over 153,000 tokens of work into a compact memory format of under 1,000 tokens.
• The project supported seamless transitions among multiple AI environments with zero degradation in performance or decision integrity, showcasing exceptional robustness in development workflows.

Architecture Overview

The architecture consists of a core engine that interfaces with various agent environments, ensuring effective memory management through:

• Adapters & Hooks: Supporting various agent actions and lifecycle events.
• Agent Provenance Engine: Tracking the origin of decisions made during sessions, ensuring that every entry is logged appropriately based on which agent was active.
• Local-First SQLite Store: All memory records are stored locally, reinforcing privacy and providing control over project data.

Multi-Agent Quickstart

Setting up engrim is straightforward. Use the following command to automatically configure the engine with detected environments:

engrim setup

Provenance Tracking

Maintaining the integrity of project memory is crucial. engrim captures the source of every memory entry, allowing for clarity in decision-making. This functionality enhances collaboration across multiple agents working on the same codebase, making it easy to track contributions and changes.

Security & Privacy

engrim prioritizes user data security:

• All memory records are stored locally with no telemetry or cloud dependencies, ensuring complete privacy.
• The SQLite database is secured with restricted file permissions to protect sensitive information from unauthorized access.

In summary, engrim offers a sophisticated and flexible solution for managing episodic memory across various AI platforms, enhancing productivity while safeguarding user data and architectural decisions.