Engineering blog
Benchmarks, architecture writeups, and decisions from our retrieval and compliance engineering. Older than the current study, and kept here because the work still stands.
Mnemosyne ships as three packages now: the retrieval engine, an MCP server for Claude Code and Cursor, and a zero-config Ollama bridge for fully local code search. Same engine, three delivery vectors, zero cloud required.
The architecture behind Mnemosyne, a six-signal hybrid retrieval engine that reduces LLM context waste by 73%. Timeline, design decisions, and how it compares to competing approaches.
We ran the same codebase question through two search pipelines, Mnemosyne semantic retrieval and standard grep/glob/read, on Claude Opus 4. Mnemosyne was 2.4x faster, 5.6x cheaper, with equivalent answer quality. Here are the full numbers.
We built Mnemosyne to stop burning tokens on irrelevant code. Two benchmarks against an 844-file production codebase: 74-78% token reduction, 92-98% faster queries, 13 tool calls reduced to 1. Zero dependencies, open source.
How we built two complementary tools, RuleBrief for regulatory monitoring and PrivacyPeep for privacy policy analysis, and the shared design principles behind both.
A technical overview of a ranked decision pipeline, how market data becomes ranked recommendations, how the system learns from outcomes, and why every decision is traceable and auditable.