KBTRANSFER

An open protocol + reference implementation for a local-first, federated knowledge base between AI agents.

Users keep their own Karpathy-style markdown wiki. Their agents both contribute to it (ingest sources, write patterns, log decisions and failures) and consume verified knowledge packs distilled from other users' wikis. Sharing is cryptographically signed, verifiable offline, and distributed through a git-based registry.

Status: Planning locked 2026-04-14. Implementation in progress. See ROADMAP.md for phased delivery.

Why it exists

Two patterns already solve half the problem each. Neither bridges to the other.

• Karpathy's LLM Wiki (intra-agent): an LLM maintains a living markdown wiki that compounds over time instead of rediscovering knowledge per query. Original gist.

• AutoEvolve Packs (inter-agent): pattern-level knowledge packs, Ed25519-signed, four-attestation, redacted by a bias-isolated pipeline, verifiable offline. Specs live under specs/current/; dogfood iteration history under reports/.

KBTRANSFER is the missing bridge: the live wiki where knowledge grows, the distillation pipeline that turns a slice of that wiki into a shareable pack, and the consumption layer that imports other people's packs without polluting your own wiki.

What makes it different

Experience is first-class. The default wiki skeleton reserves two dedicated folders, decisions/ and failure-log/, alongside the usual patterns/ and entities/. Raw reference docs can be rediscovered from public sources; what a team actually tried, what broke, and why they chose X over Y cannot. These experience pages are the pages that distill into the most valuable packs.

One protocol, three tiers. Individual developers, small teams, and regulated enterprises share the same pack format but scale redaction strictness, trust posture, and review gates to match their compliance needs. A single .kb/tier.yaml drives every tier-dependent behavior.

MCP-native. The reference implementation ships as an MCP server, so Claude Code, Cursor, Claude Desktop, and any MCP-aware agent can drive it without a custom adapter.

Plain markdown + git. User data is a folder of markdown files tracked by git. No database owns the truth. A git clone is a complete backup; a git diff is a complete audit trail.

Repository map

Licenses

• Code (reference/, tests/, examples/): Apache License 2.0.
• Specs and reports: CC BY-SA 4.0.

See LICENSE.md for details.

Attribution

Built on top of the AutoEvolve Packs specification and the LLM-Wiki pattern from Andrej Karpathy. Full background in specs/current/autoevolve-packs-overview.md.