content-addressed-storage

Steadlith reuses unchanged RAG chunks with content-defined identities, cache-aware planning, and transactional indexing.

Local content-addressed archive with observation history. Stores bytes by SHA-256, preserves locator state as contiguous spans, compresses with zstd and corpus-trained dictionaries. SQLite-backed.

Single-file, zero-dependency Python vault: write-once (WORM) storage with content-defined dedup, AES-256-GCM at rest, and 4K streaming from a Pi-sized RAM budget

a bittorrent client with content-addressed storage

Content-addressed asset management system for ingesting, deduplicating, cataloging, and searching large heterogeneous data collections across local and S3-compatible storage.

LENS-64S: 1-byte tiny-pointer CAS + suite (KnotVault, PrefixForge, DriftLedger, CheckpointTime). Stdlib Python. Measured locality.

A dataflow evaluator whose graph is an auditable record. Artifacts declare what they are fit for, and no node raises that without citing a measurement.

Decentralized, poisoning-resistant distribution for ML models and datasets. Publishers sign once, untrusted mirrors distribute, and any consumer verifies the exact signed bytes with full provenance.

fsspec backend for Ethereum Swarm: bzz:// URLs for pandas, dask, zarr and FUSE mounts, backed by a Bee node

Evidence-first visual artifact pipeline for field agents. Content-addressed vault, model-agnostic observation jobs, Merkle manifests.

Verifiable serverless SQLite hosting on Ethereum Swarm: read-only VFS for lazy SELECT + a publisher for the write path over swarmfs

Local-first solo research tool that stores claims, evidence, objections, questions, and syntheses in a Merkle-style graph on content-addressed storage.

A build system for LLM pipelines. Every step records what it actually read - files, prompt templates, model ids - so a change re-runs only the steps it can reach. Over a 30-commit replay that saved 92.7% of model calls, with all 30 results identical to a from-scratch build.

A lightweight, Git-inspired version control tool tailored for documents and large files, using content-addressed storage without the hassle of symlinks.

Explainable cache invalidation for trusted local task pipelines

Deterministic assemble and verify tooling for bounded artifact sets.

Compressed, content-addressed memory engine for RAG & LLMs — ~50% less vector storage at 96%+ retrieval quality. BLAKE3 dedup, Zstandard, binary→int8 rerank.

A persistent project-memory layer for Claude Code. Uses a content-addressed knowledge tree with Unix-style tools for querying. Claude orients to your project and your workstyle, with low overhead.

Specification-first exact reuse and safe federation for agent and workflow computations

Research platform for AI agent integrity under adversarial pressure: append-only hash-chained belief store with content-addressed artifacts and replay / regression / falsification / invariants gates.