CodeAbra/iai-personal-memory-engine

What it solves

It provides AI coding agents with a long-term, local memory system. This eliminates the need for users to manually remind the AI of previous conversations or context, as the engine automatically captures every interaction verbatim and serves relevant memory slices back to the agent during new sessions and turns.

How it works

The project runs as a local server using the Model Context Protocol (MCP). It employs a multi-tiered memory model:

• Episodic Memory: Stores timestamped, write-once fragments of conversations.
• Semantic Memory: Generates summaries from related episodes during idle periods.
• Procedural Memory: Learns behavioral parameters over time.

It uses a custom native engine written in Rust for embeddings and graph kernels, alongside a hyperdimensional memory substrate (Lilli HD) and a community-detection algorithm (MOSAIC) to organize data. The system operates locally with AES-256-GCM encryption, ensuring no telemetry or cloud accounts are required.

Who it’s for

Developers using MCP-compatible AI agents (such as Claude Code, Cursor, or Codex CLI) who want their AI assistants to have persistent, private, and high-fidelity memory of their coding workflows.

Highlights

• Local-First Privacy: Embeddings, storage, and the engine run entirely on the user's machine with no API keys or accounts required.
• Automatic Context Injection: Integrates with hosts like Claude Code to provide session-start recall and per-turn updates without manual prompting.
• High Fidelity: Focuses on literal preservation of conversations rather than just summarized approximations.
• Comprehensive Toolset: Includes 14 MCP tools for various types of recall (temporal, structural, lexical), capture, and memory management.
• Built-in Dashboard: A local interface (iai brain) to search the store, manage memories, and visualize graph neighborhoods.