How Knowledge Graduation Works: From Volatile Facts to Permanent Memory

Not all facts are created equal.

A random detail mentioned in passing shouldn't take up the same mental space as a critical architectural decision. A preference shared once shouldn't carry the same weight as a pattern that's proven useful across dozens of conversations. A temporary constraint shouldn't be remembered with the same permanence as a fundamental company policy.

Yet most systems treat all information the same way. They store everything, weight everything equally, and search through everything when you need to find something. The result is a system buried in noise.

Cortex's knowledge graduation system solves this by letting facts earn their place in memory. It's a four-tier system where facts start at the bottom and can rise through the ranks based on demonstrated usefulness. Facts that prove themselves valuable graduate to longer lifespans and higher priority. Facts that don't prove useful fade away.

The Four Tiers: From Volatile to Permanent

Let's walk through each tier of knowledge graduation:

Tier 1: Volatile Memory (1 Hour)

When a fact first enters the system, it starts as volatile memory. This is the working memory for a single conversation or work session. A volatile fact persists for one hour and requires a usefulness score of 0.40 or higher to survive.

Think of volatile memory like a notepad you're working with right now. It holds information that might be useful in the immediate term, but you're not yet sure if it matters. The threshold for staying in volatile memory is low because you're still exploring.

A volatile fact might be:

• A parameter value mentioned during a problem-solving session
• A temporary constraint for a specific project
• A user preference shared for the first time
• A decision rationale that was just explained

The one-hour window is intentional. It's long enough to be useful within a single work session but short enough that facts don't accumulate endlessly. After one hour, a fact either moves up to the next tier or gets discarded.

Tier 2: Daily Memory (24 Hours)

Facts that prove useful graduate to daily memory. They stick around for 24 hours and require a usefulness score of 0.60 or higher. This is where facts that show genuine relevance are kept.

Daily memory captures information that's proven useful once or twice. It's the working knowledge that you reference regularly but haven't yet become permanent institutional knowledge. It's the "we tried this approach before and it worked" or "this person prefers communication this way" type of information.

A fact might graduate to daily memory when:

• A user preference has been referenced in multiple conversations within a day
• A technical constraint is confirmed across different contexts
• A decision pattern emerges from two or more interactions
• A solution that worked once is relevant again

Daily memory requires a higher usefulness score (0.60) because the system is being more selective. Not every fact that worked once deserves to stick around longer. The system wants to see demonstrated relevance.

Tier 3: Stable Memory (7 Days)

The third tier is stable memory. Facts here persist for seven days. To reach stable memory, a fact must:

• Achieve a usefulness score of 0.80 or higher
• Be accessed and used at least three times
• Persist for at least two days of previous use

Stable memory is where institutional patterns live. These are facts that have proven themselves useful repeatedly and consistently. They're the established practices, confirmed preferences, and validated approaches.

Examples of stable memory:

• Your company's standardized deployment process
• Key architectural patterns that work well in your codebase
• Team members' communication preferences and work styles
• Recurring technical constraints that shape decisions
• Project dependencies and timing that affect planning

The higher threshold (0.80) and the multi-factor requirements (score plus access count plus temporal persistence) mean only genuinely useful facts reach stable memory. But once they do, they're available for a week without needing to prove themselves again.

Tier 4: Permanent Memory (Forever)

The final tier is permanent memory. Facts here stay forever. They only reach permanent status after graduating through the entire system with sustained usefulness.

Permanent memory holds the most critical knowledge:

• Core company policies and principles
• Fundamental architectural decisions and their rationales
• Key team structures and organizational hierarchies
• Established security requirements and compliance constraints
• Historical decisions that shaped company strategy

To reach permanent memory, a fact doesn't just need high scores. It needs to demonstrate consistency and importance over time. The system is extremely conservative here because permanent memory is the foundation of institutional knowledge.

The Scoring Formula: How Facts Earn Their Promotions

Behind knowledge graduation is a simple but powerful scoring system. Every fact in memory gets a usefulness score based on three factors:

Usefulness Score = (0.40 × Usefulness) + (0.30 × Efficiency) + (0.30 × Confidence)

Let's break down each component:

Usefulness (40% weight): Did the injected fact appear in the response? This is measured through keyword overlap analysis. If a fact was retrieved and injected into a conversation, the system checks whether that fact's key terms appear in the AI agent's response. If they do, it's counted as useful. If the injected fact was ignored, it gets a lower usefulness score.

Efficiency (30% weight): How often was the fact used when it was retrieved? This is calculated as the ratio of times a fact was used divided by times it was injected. A fact that's injected five times and used in all five responses has an efficiency score of 1.0. A fact injected ten times but used only five times has an efficiency score of 0.5. The system rewards facts that are consistently relevant.

Confidence (30% weight): How certain is the system about this fact's validity and usefulness? This can be influenced by the source (did it come from multiple agents confirming the same thing?), the temporal consistency (has this fact remained true over time?), and explicit user feedback. Higher confidence means the fact is more likely to be genuinely valuable.

This formula is deliberately balanced. A fact with high usefulness but low efficiency won't graduate. A fact with high efficiency but low usefulness won't graduate. The system requires demonstrated, consistent value.

A Concrete Example: One Fact's Journey

Let's follow a real fact through the graduation system to see how this works in practice.

Day 1, Hour 1: A Fact is Born During a code review conversation, a developer mentions: "We always use fastcopy for bulk operations because it's three times faster than the standard approach, and we've had memory issues with large transfers otherwise."

This fact enters volatile memory. It has a usefulness score of 0.35 (just below the 0.40 threshold but present in the system). The one-hour timer starts.

Day 1, Hour 0:45 The developer asks a follow-up question about optimizing a data pipeline. The memory system injects the fastcopy fact. The agent incorporates it into the response, recommending fastcopy for bulk operations.

The keyword overlap analysis detects that "fastcopy" appeared in both the injected fact and the response. The fact's usefulness score jumps to 0.58, and it's marked as validated. The fact graduates to daily memory.

Day 2, Hour 10 A different developer asks about performance improvements for large data transfers. The memory system injects the fastcopy fact again. Again, it appears in the response with detailed reasoning about why it's the right choice.

The fact's access count is now 2 and injection count is 2 (efficiency: 1.0). The temporal consistency is solid. The usefulness score reaches 0.72. The fact remains in daily memory.

Day 3, Hour 2 An AI agent assisting with infrastructure improvements encounters the fastcopy recommendation again. This is the third injection and third use. The fact has persisted for two days. The usefulness score reaches 0.84.

All conditions are met: score >= 0.80, access count >= 3, and temporal persistence >= 2 days. The fact graduates to stable memory. It will now persist for seven days without needing additional validation.

Day 4, Day 5, Day 6, Day 7, Day 8 Over the next week, the fastcopy fact is injected eight more times across different conversations. In six of those cases, it appears in the response. The efficiency remains high (14/11 = 1.27, capped at 1.0). The confidence grows as multiple different team members and AI agents reference the same best practice.

Week 2 After demonstrating consistent usefulness over two weeks, appearing across multiple team members and contexts, and maintaining high scores, the fastcopy fact is evaluated for permanent memory. It clears the bar. The system recognizes this as institutional knowledge that should be preserved indefinitely.

From this point forward, whenever anyone asks about bulk data operations or performance optimization, the fastcopy best practice is available.

The Contrast: Why Cortex Beats Flat Memory Systems

OpenClaw, an alternative system, uses a file-first approach. Every piece of information gets stored in Markdown files. The system uses SQLite with full-text search to retrieve information. It's straightforward and simple.

But here's the problem: all history is treated equally.

In OpenClaw's system, a throwaway comment about the weather has the same retrieval weight as a critical architectural decision. A decision mentioned once is just as prominent as a decision confirmed across ten conversations. Nine months of outdated information is just as searchable as yesterday's decision.

The result: 90% noise, 5% operational knowledge, 5% institutional knowledge.

Cortex's graduation system filters this ruthlessly. Information that matters rises to the top. Information that doesn't fade away. The signal-to-noise ratio improves every single day.

More importantly, Cortex's system learns. It doesn't just store facts; it learns which facts matter most. Over time, your knowledge base becomes increasingly refined, increasingly focused on what actually drives value for your specific context.

That's the difference between having a memory and having a learning system.

Knowledge graduation isn't just a storage mechanism. It's how AI agents become genuinely intelligent about your context, your constraints, and your goals. It's how they stop being replaceable search indexes and become actual team members who learn from experience.

See how knowledge graduation works in your organization. Sign up at launchcortex.ai for a free 14-day trial and watch your agents learn and improve from day one.