How persistent AI memory transforms GitHub repositories into living docs

In software development, the most valuable knowledge often disappears with the engineers who wrote it. Commit messages document what changed, but rarely capture why decisions were made. A new tool called Shadow CTO is changing that by installing a persistent AI memory system directly into GitHub repositories.

This open-source solution watches every commit, pull request, and issue in real time, extracting engineering decisions and their rationales. Instead of generating summaries from static documentation, it provides precise answers based on actual commit messages, pull request discussions, and historical context. The result is a living institutional memory that never quits, never forgets, and remains accessible to the entire engineering team.

Turning Git commits into institutional knowledge

Traditional documentation approaches rely on developers manually recording the reasoning behind technical choices. Over time, this institutional knowledge erodes as team members leave and priorities shift. Shadow CTO automates this process by treating each GitHub repository as a distinct memory container.

Every repository gets its own dedicated Hermes Agent session, identified by a unique X-Hermes-Session-Id header. This persistent memory persists across multiple conversations and months of development activity. When engineers ask questions about past decisions, the system responds with exact commit references, pull request context, and the original problem statements that drove changes.

For example, asking "Why was Redis removed from the authentication system?" returns a detailed response including the specific pull request number, the date of implementation, the production incident that triggered the change, and the tradeoffs considered. This level of precision eliminates the need for developers to comb through git history manually or rely on potentially outdated documentation.

Real-time decision tracking and failure pattern detection

Shadow CTO doesn’t just store decisions—it actively analyzes them to identify emerging patterns and recurring issues. The system automatically extracts meaningful engineering decisions from every commit and pull request, classifying them by type and assigning confidence scores based on the quality of reasoning provided.

The dashboard presents these decisions in an organized interface where teams can browse historical choices, understand the evolution of technical approaches, and spot contradictions or reversals in strategy. This historical perspective helps prevent repeated mistakes and enables better technical planning.

One of the most powerful features is the autonomous failure pattern detection. Every night at 2 AM, Shadow CTO registers a scheduled job with Hermes Agent to analyze the past 30 days of engineering decisions. The system identifies components that consistently break, decisions that get frequently reversed, or technical debt that accumulates over time. These insights appear automatically in the dedicated Patterns tab, providing teams with actionable intelligence without manual intervention.

Technical architecture: A persistent memory stack

Shadow CTO’s effectiveness stems from its architectural decisions, particularly its use of Hermes Agent as both memory storage and scheduling backbone. The system operates across four distinct layers that work together seamlessly.

Memory isolation per repository

Each GitHub repository maintains its own isolated memory session within Hermes Agent. This architectural choice prevents knowledge bleeding between unrelated projects while ensuring contextual continuity within individual codebases. The system uses repository-specific session IDs to maintain these separate memory spaces, effectively giving each project its own persistent AI brain.

# hermes_client.py
response = await self._openai.chat.completions.create(
    model=self.model,
    messages=messages,
    extra_headers={"X-Hermes-Session-Id": session_id},  # Per-repo memory isolation
)

Structured decision extraction at ingest time

When new commits or pull requests arrive, they’re immediately processed through a structured prompt that asks Hermes to determine if a meaningful engineering decision was made. If so, the system extracts the rationale, classifies the decision type, and assigns a confidence score based on the clarity of the reasoning.

# services/ingestion.py
INGEST_SYSTEM_PROMPT = """
You are the Shadow CTO for {repo_name}. When you receive a commit, PR, or issue, 
determine if a meaningful engineering decision was made, extract the rationale, 
and classify the decision type. Respond in JSON.
"""

This structured extraction happens in real time as new code changes are processed, ensuring that institutional knowledge is captured immediately rather than retroactively.

Streaming natural language responses

Shadow CTO implements Server-Sent Events to stream responses from Hermes memory directly to the frontend interface. This approach provides users with progressively revealed answers rather than waiting for complete processing, creating a more interactive experience similar to conversational AI assistants.

# routers/query.py
event_stream = []
async for chunk in hermes.stream_chat(
    messages=messages,
    session_id=repo.hermes_session_id,  # Query the specific repo's memory
):
    yield f"data: {chunk}\n\n"

Autonomous job scheduling integration

The system registers two automated jobs directly with Hermes Agent’s scheduler during startup. These jobs run completely autonomously, analyzing historical decisions and identifying patterns without any human prompting. The scheduled analysis appears in the Hermes jobs dashboard, making the autonomous behavior transparent to users.

# jobs/cron_setup.py
await hermes.create_job(
    name="shadow-cto-daily-patterns",
    schedule="0 2 * * *",
    prompt=(
        "You are the Shadow CTO. Review the engineering decisions stored "
        "in your memory from the past 30 days. Identify recurring failure "
        "patterns—components that keep breaking, decisions that get reversed, "
        "or technical debt accumulating."
    ),
)

Why persistent memory beats traditional documentation

Most code documentation systems rely on static files or vector databases that retrieve information based on keyword matching. Shadow CTO takes a fundamentally different approach by using persistent agent memory that accumulates understanding over time.

Traditional RAG systems answer questions about what exists in documents, while Hermes-based memory answers questions about what the AI has learned from observing the repository. This distinction means Shadow CTO can provide context-aware responses that understand the evolution of technical decisions rather than just retrieving isolated fragments.

The system also eliminates the need for separate analysis pipelines. By integrating directly with Hermes Agent’s scheduling system, Shadow CTO performs autonomous analysis as part of the same memory infrastructure, reducing complexity and ensuring consistency between ingested data and analytical outputs.

Looking ahead: AI memory as engineering infrastructure

Shadow CTO demonstrates that persistent AI memory isn’t just a demonstration feature—it’s becoming essential infrastructure for modern software development. By capturing the reasoning behind engineering decisions in real time and making this knowledge continuously available, teams can reduce onboarding time, prevent repeated mistakes, and maintain institutional knowledge across personnel changes.

As AI agents become more integrated into development workflows, systems like Shadow CTO will likely evolve from optional tools to standard infrastructure components. The ability to query historical reasoning behind every technical decision may soon become as fundamental as version control itself, ensuring that the most valuable institutional knowledge—the reasoning behind code—remains accessible to all engineers, present and future.