Shodh Memory

Shodh Memory solves a narrow but real problem: agents forget across sessions, and the common fix — an LLM in the memory loop — is slow, expensive, and non-deterministic. Shodh routes around that entirely. Storing, consolidating, and recalling memories never touch a model; learning happens through Hebbian edge updates and decay curves, so the same input always produces the same memory and each memory op costs $0. The strongest evidence for its engineering is in the retrieval layer. Instead of shipping vector search with a wrapper, it runs a Vamana vector index, a Hebbian knowledge graph, and a temporal decay index side by side, then ranks across all three. That is what makes causal retrieval through backward graph walks possible — the blog post on 2026-06-13 frames this as 'the memory problem vector search can't solve' — and it also enables blind-spot, shallow-knowledge, orphaned-cluster, and stale-zone detection, i.e. surfacing what an agent has forgotten before it matters. Practical strengths: it installs in one command (claude mcp add shodh-memory -- npx -y @shodh/memory-mcp), runs on Raspberry Pi Zero through air-gapped industrial PCs, and exposes 37–45 MCP tools plus a REST API with a /health endpoint, auth, rate limiting, and CORS. The weaknesses are honest ones. It is local-first, so there is no built-in cross-machine sync and no team-shared memory out of the box; setup is manual per environment. Because it deliberately keeps LLMs out of the loop, it does not generate LLM-based summaries or semantic enrichment during storage — if your workflow expects an LLM to summarize and label memories on write, this is the wrong architecture. Working memory capacities are intentionally small (sensory buffer ~7 items, working memory ~4 chunks) as a model of human memory, not a configurable budget. And because models auto-download on first use (~37MB to ~/.cache/shodh-memory/), the very first run needs a network connection even though steady-state operation is fully offline. Where it fits: edge and IoT deployments, ROS2/Zenoh robotics that need offline recall of learned paths and object associations, fault-tracing agents that recall past decisions and trace causes, and privacy-first teams with no cloud dependency. Where it doesn't: teams wanting a hosted memory SaaS with SSO, RBAC, and a management UI, or non-technical users who want a GUI-only tool rather than a binary and an API.