Your agent keeps making the same mistakes. This repo remembers.

What it does

MisakaNet stores debugging failures as Markdown files in a git repository, graded by evidence level from raw intake to production-proven. Agents query the corpus via an MCP server or a public HTTP endpoint, retrieving past fixes with BM25 search over the Python standard library. If no lesson matches, the system returns an honest no_match and can turn the dead end into a new intake for future agents.

The interesting bit

The project treats git as the database: lessons are DCO-signed commits that diff, fork, and review like code, while the retriever deliberately uses zero-dependency BM25 instead of embeddings so it runs air-gapped without model rot. Evidence is graded E0–E4 rather than asserted as truth, which lets an agent weigh a community rumor differently from a battle-tested fix.

Key highlights

• 411+ indexed failure-recovery lessons across domains like Docker, CI, WSL, and MCP itself
• Zero-dependency BM25 retrieval that works offline with no vector database or embedding model
• Evidence-level grading (E0 intake through E4 production reuse) instead of binary “verified” flags
• MCP server, WebMCP, llms.txt, and A2A discovery interfaces for agent-native access
• Published benchmarks on Cloudflare Workers AI show lesson context doubles hit rates for small models

Caveats

• BM25 matches words, not meaning, so paraphrased failures the corpus has never seen will miss; the README explicitly calls this a designed trade-off, not a solvable bug
• The majority of lessons sit at evidence levels E0–E2, meaning the library leans more toward “reported failures” than “production-proven fixes”
• Package naming across npm and PyPI is tricky enough that the README dedicates a section to a “two-package trap” caused by confusing the installer, plugin, library, and server distributions

Verdict

Worth a look if you run AI agents that repeatedly hit the same environment-specific errors and you want a searchable, auditable memory without spinning up a vector database. Skip it if you need a general-purpose skill store or semantic retrieval that understands rephrased symptoms.