// the take
Shodh-Memory is a Rust binary that gives AI agents (and robots) persistent memory without any cloud dependency or LLM API calls. It stores memories in a three-tier RocksDB structure — working, session, long-term — with Hebbian reinforcement and mathematical decay borrowed from neuroscience literature. Ships as a single binary with MCP, REST, Python/npm/Rust SDKs, and a Zenoh transport layer for ROS2 robotics.
- Zero LLM calls on the hot path is a real differentiator. Competitors like mem0 and Cognee burn 2–3 OpenAI calls per store; Shodh uses local MiniLM embeddings and algorithmic scoring, which is why the 55ms store latency is plausible and the 763ns entity lookup is believable.
- The neuroscience grounding is not decorative. Cowan's working memory model and Wixted's decay functions are directly reflected in the three-tier architecture — memories don't just accumulate forever, they actually fade or strengthen based on usage frequency, which solves a real retrieval-quality problem.
- The robotics integration is genuinely novel — 26 mission-specific payload fields (geo_location, heading, reward, terrain_type, etc.), haversine spatial recall, and first-class Zenoh transport mean a Spot or PX4 drone can query 'what happened near this GPS coordinate last mission' without any cloud round-trip.
- Actually shipped to every major registry: crates.io, PyPI, npm, Docker Hub, MCP Registry, Homebrew — which suggests the release pipeline is real, not vaporware.
- The repo has junk committed to it: `mcp-publisher.exe`, `full_log.txt`, `job_log.txt`, `bhashini-authorization-letter.html`, `STPI_PITCH_DECK.md`, `Shodh_memory_IITD.pdf`. These are files from a grant application or pitch process. A project disciplined enough to implement DiskANN-variant indexes should know not to commit pitch decks to its source tree.
- The Python and npm SDKs are thin HTTP wrappers, not native bindings — users still need the Rust binary running as a sidecar. The README buries this; someone reading `pip install shodh-memory` will assume they get a self-contained library.
- The custom vector DB (`src/vector_db/vamana.rs`, `spann.rs`, `pq.rs`) reimplements DiskANN/SPANN in a side project. These are complex algorithms with subtle correctness requirements; the risk that this implementation has silent recall quality bugs is non-trivial at 217 stars with no published recall accuracy numbers against a standard benchmark.
- No schema migration story. The RocksDB storage format is undocumented and the codebase has a `bugs-remedies/` directory with dated audit files suggesting the storage layer has already had correctness issues. Upgrading the binary without losing your memory data is an open question.
We read the README, the directory tree and a key source file in June 2026. Nobody pays to be listed. How we judge →