Shodh Memory

Shodh Memory takes a genuinely different path from mem0, Zep, or cognee by never calling an LLM in the memory loop. Instead, it uses Hebbian learning and decay curves — algorithms that run in microseconds and produce deterministic results. That means no surprise bills, no latency spikes, and no privacy trade-offs: every memory stays on your machine in a ~30MB Rust binary. We'd reach for Shodh when building autonomous agents that need persistent, ground-truth memory without reliance on cloud APIs or LLMs. The causal retrieval feature — highlighted in recent blog posts — reconstructs chains of reasoning, answering "why did I decide X?" in a way vector memory cannot. That's a real edge for debugging agent behavior or building explainable systems. Where it bites: there is no native cloud sync or team-sharing. If your use case demands collaborative memory across devices or a hosted management UI, you'll need to build that yourself. Enterprise features like SSO and RBAC are not available yet. And non-technical users won't find a GUI — it's primarily CLI and API. Compared to mem0, which also offers open-source memory but leans on LLM calls for summarization and enrichment, Shodh is faster and cheaper per operation (literally $0). However, mem0 provides a cloud-hosted option and a higher-level API for less technical teams. Similarly, Zep and cognee offer cloud sync and team features but charge monthly fees. In practice, Shodh shines on edge devices: Raspberry Pi, Jetson, industrial PCs. It's also a strong choice for robotics workloads via ROS2/Zenoh. The published benchmarks (<1μs graph lookups, 34-58ms semantic search) are backed by 1,089 tests. Just don't expect a polished frontend or enterprise support — this is a developer's tool first.