Memory systems became infrastructure the moment your agent started making decisions from stale context

I thought persistent memory would make our agents smarter. Instead, it made them slower, more expensive, and weirdly nostalgic.

Our support agent started referencing customer complaints from six months ago in current conversations. Our coding agent kept suggesting deprecated libraries because they were mentioned in old project files. The marketing agent was writing copy in the style of campaigns we'd abandoned.

The problem wasn't the memory system — it was that we treated memory like a database when we should have been treating it like infrastructure.

Memory isn't just storage. It's the context your agent uses to make every decision.

Here's what we learned building memory that actually scales:

Memory systems need garbage collection, not just garbage in/garbage out prevention.

We built a three-tier system:

• Working memory — Current session context, expires after 24 hours
• Knowledge graph — Facts, relationships, and procedures that persist
• Archive — Historical context, searchable but not loaded by default

The key insight: your agent doesn't need to remember everything. It needs to remember the right things at the right time.

Working memory handles the immediate context — what we're working on right now, recent decisions, current project state. This gets wiped clean every day unless explicitly promoted.

The knowledge graph stores the permanent stuff — customer preferences, system configurations, process documentation. This is what makes your agent actually learn from experience instead of starting from scratch every conversation.

The archive is where old conversations go to die gracefully. Still searchable when needed, but not cluttering up every decision.

The promotion rules are where the magic happens:

// Promote to knowledge graph if:
- Referenced 3+ times in different sessions
- Marked as "remember this" by user
- Involves system configuration or process
- Corrects a previous mistake

// Archive if:
- Older than 30 days and not referenced
- Marked as resolved/completed
- Superseded by newer information

We also built decay functions. Customer preferences from last week are weighted higher than ones from last year. Recent project decisions override old ones unless explicitly flagged as permanent.

The result? Our agents got faster (smaller active memory), cheaper (less context to process), and more accurate (relevant context instead of everything-context).

Memory systems are infrastructure now. Your agent's decision quality depends on memory hygiene, not memory size.

Stop treating agent memory like a append-only log. Start treating it like the foundation of everything your agent knows.