Slmfs — AI skill for Claude Code

SLMFS

[](https://opensource.org/licenses/Apache-2.0) [](https://isocpp.org/) [](https://www.python.org/)

**Giving AI memory the laws of physics and geometry.**

SLMFS (Superlocal Memory FileSystem) is a zero-copy, mathematically rigorous long-term memory engine for autonomous AI agents (like Claude Code). Instead of relying on ad-hoc RAG pipelines, hardcoded forgetting thresholds, or simple cosine similarities, it grounds agent memory in **Information Geometry**, **Algebraic Topology**, and **Stochastic Differential Equations (SDE)**.

To the AI agent, it looks like a standard, transparent Markdown file system. Under the hood, it's a high-performance C++23 engine running a continuous physical simulation of memory.

Theoretical Origin

This implementation is based on the visionary concepts described at **[superlocalmemory.com](https://www.superlocalmemory.com/)**.

The Mathematical Foundations

• Fisher-Rao Retrieval Metric (Information Geometry): Memories are not points, but probability distributions (diagonal Gaussian families). Retrieval scoring is derived from the Fisher information structure, smoothly transitioning from linear cosine similarity to geodesic distance over the first few accesses.
• Sheaf Cohomology for Consistency (Algebraic Topology): Ad-hoc contradiction checks are replaced with algebraic guarantees. By computing the coboundary operator over the local memory graph, the system mathematically detects logical contradictions and gracefully self-corrects the state.
• Riemannian Langevin Lifecycle (Non-Euclidean SDE): Memory nodes exist on a Poincare disk. Frequently accessed memories are pulled to the center. Neglected memories undergo Langevin diffusion, naturally drifting toward the hyperbolic boundary where space compresses. There are no "delete a