knowing 0.10.0

Content-addressed knowledge graph for software systems

knowing

A self-adapting code intelligence engine that gets smarter with scale, not dumber.

Content-addressed graph of code relationships with density-adaptive retrieval. Observes its own graph structure and adjusts how it searches: on small repos it finds symbols by keyword; on large enterprise codebases it automatically shifts to structural navigation. 34 edge types, cryptographic proofs, hierarchical Merkle snapshots.

This is the Python wrapper package that downloads and runs the knowing binary.

Install

pip install knowing

Usage

# Register and index a repository
knowing add ./path/to/repo

# Graph-ranked context for an agent task (84% fewer tokens than JSON)
knowing context -task "refactor auth" -format gcf

# Find affected tests for changed files
knowing test-scope -files internal/auth/session.go

# Explain why a symbol ranked where it did
knowing why -task "refactor auth" -symbol "SessionHandler"

# Run the MCP server with live file watching
knowing mcp --watch

# Verify graph integrity
knowing fsck

What It Does

knowing indexes code across 12 languages and 26 extractor packages into a content-addressed knowledge graph. Every node, edge, and snapshot is SHA-256 hashed with domain-type prefixes. Snapshots are structured as hierarchical Merkle trees (repo -> package -> edge-type -> leaf), enabling O(packages) diffs instead of O(edges) full scans.

For AI agents: 28 MCP tools + 8 MCP resources serve graph-ranked context over stdio or HTTP. The GCF wire format delivers 84% token savings versus JSON. Agents get trustworthy, cacheable, replayable context with provenance and confidence on every edge.

Key capabilities:

• 26 extractor packages: Go, TypeScript, Python, Rust, Java, C#, Protobuf, Terraform, SQL, Kubernetes, CloudFormation, Docker Compose, GitHub Actions, Helm, GraphQL, and more
• Hierarchical Merkle diffs: 216x faster on real graphs (~24.9K edges), 517x at 100K edges
• Subgraph cache: 93x faster repeat queries via content-addressed cache keys
• Runtime fusion: OpenTelemetry trace ingestion merges static and runtime views
• Graph notes: general-purpose metadata layer for community assignments, context pack persistence, and feedback annotations
• knowing fsck: git-style integrity verification of the entire graph
• Feedback loop: rankings improve with use as agents mark useful symbols

Agent Integration

Add to .mcp.json:

{
  "mcpServers": {
    "knowing": {
      "command": "knowing",
      "args": ["mcp", "--watch"],
      "transport": "stdio"
    }
  }
}

Documentation

Full docs, architecture, benchmarks, and roadmap at https://blackwell-systems.github.io/knowing

Source: https://github.com/blackwell-systems/knowing