arc-context 1.0.0

ARC (Agent Reasoning Context) — portable, verifiable context packaging for AI agents

ARC - Agent Reasoning Context

AI answers are hard to trust. ARC makes them inspectable.

ARC packages repository knowledge into a verifiable context artifact, so agents can load precise, traceable context instead of piecing it together on the fly.

ARC doesn't make AI smarter. It makes it inspectable.

---

Why ARC

AI agents don't operate on a stable, structured view of your system.

They piece together context on the fly from:

• conversation history
• files they read
• retrieval results

This works - but the result is often:

• redundant
• noisy
• hard to verify

ARC changes that.

Instead of guessing and searching blindly, agents start from a pre-built, structured context with evidence.

guess -> search -> hope

becomes:

load -> locate -> prove

---

How it works

repo + docs + tickets + policies
  -> arc build .
  -> semantic layers + manifest
  -> project.arc
  -> agent loads only what it needs

ARC builds a structured representation of your system once, then lets agents query it efficiently and reliably.

---

Install

pip install arc-context

---

Quick start

arc build . --out project.arc                         # extract claims, decisions, evidence from source
arc inspect project.arc                               # show layers, blob count, manifest summary
arc verify project.arc                                # check Merkle integrity, detect tampering
arc load project.arc --task "review this auth change" # retrieve only context relevant to the task
arc diff project-v1.arc project-v2.arc                # compare two archive versions

---

What ARC does

• Builds structured context - claims, decisions, evidence pointers
• Makes answers inspectable - every claim is traceable to source
• Loads selectively - hybrid vector + keyword retrieval
• Stores content-addressed - Merkle integrity, reproducible builds
• Diffs versions - track how context changes over time
• Verifies offline - no runtime dependency on external services

---

What ARC is not

• Not a vector database wrapper
• Not a runtime memory system
• Not a zip file with markdown
• Not framework-specific

---

Architecture

ARC is built as a simple 5-layer system:

Source -> Builder -> Artifact -> Loader -> Runtime

• Builder extracts semantic structure (claims, decisions, evidence)
• Artifact stores it as a content-addressed archive
• Loader retrieves only relevant context for a task

Full system design: docs/architecture.md

---

Benchmark

ARC preserves near-hybrid retrieval quality while making every result traceable and debuggable.

hybrid_arc ~ hybrid recall, but with full traceability (1.0 vs 0.0)

30 tasks per repo, 6 categories. Context recall = fraction of required facts found.

Full analysis: docs/benchmark-fastapi-vs-django.md

• ~93% of hybrid recall
• every result traceable to source file + line

---

Example

Query:

How does authentication work?

ARC returns:

• Answer
• Evidence (files + lines)
• Suggested files to inspect

---

When to use ARC

• working with large repos
• debugging AI-generated answers
• reviewing code changes with context
• building AI agents that need reliable context

---

Development

make test
make lint
make benchmark-smoke

---

License

Apache 2.0