da7-tech/mind

Mind is a memory system for AI coding agents that mimics how human memory works. It is a single Python file with no external dependencies and no API keys required. Everything runs offline on your machine. You install it by downloading one file, initializing a local memory store, and from that point forward your agent can remember and recall facts across sessions. The memory system has three layers, described in brain terms. Working memory holds the hottest, most recently used memories and is automatically injected into the rule files your agent reads at the start of each session (CLAUDE.md for Claude Code, AGENTS.md for Codex and Cursor, GEMINI.md for Gemini). The hippocampus layer is a weighted graph of connected concepts stored as a JSON file, searched using spreading activation across up to three hops rather than flat keyword matching. The cortex layer holds durable, consolidated knowledge that gets promoted from the graph over time. Between sessions you run a dream command that performs three passes: a light phase clears session markers, a deep phase applies the Ebbinghaus forgetting curve so unused memories fade and frequently recalled ones become more stable, and a REM phase clusters related memories together and promotes recurring themes into long-term cortex storage. The dream cycle is fully deterministic and produces no AI calls or token costs. The project has been extensively benchmarked. It reports 100% recall at rank 1 on graphs of 100 and 1,000 nodes with median latency under 3 milliseconds. The same recall figure holds across 10 languages in CI testing, with specific engineering for English and Arabic. A 180-day simulated soak test confirmed that core facts survived while junk notes were pruned correctly. The tool is engineered for EN + AR but has been measured on French, German, Spanish, Russian, Turkish, Chinese, Japanese, and Korean. The MIT-licensed project provides a suite of 165 tests including fuzz testing, mutation testing, and concurrency regression tests.