CodeAbra/iai-personal-memory-engine

What it solves

Most AI coding assistants forget the context of previous sessions, forcing users to repeat instructions, re-explain project orientations, or re-read files. iai-pme provides a persistent, local long-term memory that automatically captures conversation turns and injects relevant history into new sessions, reducing token costs and eliminating the need for manual "remember this" commands.

How it works

It operates as a local server using the Model Context Protocol (MCP), integrating with assistants like Claude Code, Cursor, and Zed. The system uses a custom Rust-native engine for storage, community-detection algorithms, and a hyperdimensional memory substrate to manage data.

Key mechanisms include:

• Ambient Capture: Shell hooks record every prompt and response verbatim to a local buffer.
• Automatic Recall: Relevant memory slices are injected into the conversation prefix at session start or as "foresight packs" before turns.
• Idle Consolidation: While the system is idle, it merges duplicates, strengthens frequently retrieved pathways, and prunes weak edges.
• Local-First: All embeddings and storage are handled on the user's machine with AES-256-GCM encryption, ensuring no telemetry or cloud accounts are required.

Who it’s for

Developers using MCP-compatible AI coding assistants (such as Claude Code, Cursor, Aider, or Continue.dev) who want their agent to maintain a consistent, long-term understanding of their preferences, project history, and previous discussions without relying on cloud-based memory services.

Highlights

• Token Efficiency: Claims to be ~88% cheaper than displacing agent searches by injecting concise memory packs.
• Verbatim Storage: Stores exact conversation turns rather than LLM-generated summaries, preserving historical wording.
• Rust-Native Engine: Uses a custom-built storage engine and graph kernels for high performance.
• MCP Compatible: Works with any assistant that speaks MCP-over-stdio.
• Visual Dashboard: Includes a "brain" dashboard (via iai brain) to visualize and search the memory graph.
• Privacy-Centric: Fully local execution with no API keys or telemetry.