Iris – pure-Swift ARM64 disassembler with a semantic layer
Pure-Swift ARM64 disassembler with semantic layer validated against LLVM.
Single-file memory layer for AI agents, sub mili-second RAG on Apple Silicon. Metal Optimized On-Device. No Server. No API. One File. Pure Swift
Exports a one-file 'brain' and a tiny MemoryOrchestrator API (remember/recall) so you can ditch Docker and hosted vector DBs — token-budgeted, deterministic recall and kill-9-safe durability are concrete wins. The Metal-accelerated vector search plus SQLite FTS5 fallback shows real engineering heft, but it's clearly tuned for the Apple ecosystem and the author is still asking for retrieval/eval feedback.
iOS/macOS app developers and mobile/edge AI engineers building on-device agents or apps that need private, offline RAG-style memory
Goal: replace the typical “RAG stack” with something you can ship inside an app (offline, deterministic token budgeting, reproducible retrieval).
Feedback welcome — especially on retrieval quality + eval methodology.
Pure-Swift ARM64 disassembler with semantic layer validated against LLVM.
SQLite-backed agent memory with graph viz when Mem0 and Zep already dominate.
Local RAG for Claude when Cursor and other editors already handle context.
Structured memory layers for agents—but vector search already solves this problem.
Skip SQL entirely; direct B-tree KV API cuts RocksDB memory from 121MB to 11MB.
Beats Mem0 on strict keyword matching by preserving source language instead of paraphrasing.