Your agent needs a context refresh policy — here's how to build one that prevents drift

Your agent starts sharp. Day one, it remembers your preferences, follows your patterns, knows your codebase. By week three, it's asking questions it already answered, suggesting tools you explicitly rejected, and writing code in a style you trained it out of months ago.

This isn't memory failure — it's context drift. And it kills more production agents than any other single issue.

The problem: Context windows fill with stale information

Your agent's context window is finite. As conversations get longer and memory files grow larger, the important stuff gets pushed out by the routine stuff. Your agent stops seeing the foundational context that made it useful in the first place.

Here's what gets lost first:

• Your communication preferences ("be direct, skip the pleasantries")
• Your architectural decisions ("we use Tailwind, not styled-components")
• Your workflow patterns ("always run tests before committing")
• Your quality standards ("no console.logs in production code")

Meanwhile, your context fills with yesterday's debug session, last week's meeting notes, and random API documentation that's no longer relevant.

The solution: Build a context refresh routine

Every Sunday at 6 AM, my agents run a context refresh routine. It's not complex — just three steps that reset their working memory:

## Weekly Context Refresh Protocol

1. Archive conversations older than 7 days
2. Reload core context files (preferences, standards, workflows)
3. Summarize active projects and carry forward only actionable items
4. Clear temporary files and debug artifacts
5. Reset working directory to clean state

The key insight: Your agent needs to forget the right things. Not everything is worth remembering. Most conversations are transactional. Most debug sessions are one-time events. Most meeting notes become irrelevant within days.

What to preserve vs. what to purge

Preserve:

• Your core preferences and communication style
• Architectural decisions and coding standards
• Active project context and open tasks
• Recent learnings about your codebase or business

Purge:

• Completed tasks and closed conversations
• Debug artifacts and temporary solutions
• Outdated project information
• Routine status updates and check-ins

The implementation

Add this to your agent's weekly routine:

Every Sunday at 6 AM:
1. Run context_refresh.py
2. Reload core_context.md
3. Archive old conversation files
4. Generate fresh project summary
5. Clear /tmp and /debug directories

I've been running this pattern for six months. My agents stay consistent. They remember what matters. They don't drift into generic ChatGPT behavior.

The difference is night and day. Week one agents that still act like week one agents, even months later.

Context drift is silent but deadly. Your agent won't tell you it's happening. You'll just notice it getting less useful over time. Build the refresh routine before you need it.

This is exactly the kind of operational discipline that separates toy agents from production agents. If you're running agents in production, you need patterns like this baked into your setup from day one.