Phi-Scored Persistence: Your AI Should Forget

What Is Phi-Scored Persistence?

Phi-scored persistence is a memory management strategy for AI systems where each interaction is scored using an adapted version of Integrated Information Theory's Phi (Φ) metric, and only interactions that exceed a quality threshold are persisted to long-term memory.

The principle: your AI should forget most of what happens. Not because storage is expensive, but because undifferentiated memory creates noise that degrades retrieval quality. A memory system that stores everything is a memory system that retrieves nothing well.

Phi-scored persistence is the inverse of traditional logging. Instead of "store everything, search later," it's "evaluate now, store what matters."

The Problem With Total Recall

Most AI memory systems fall into one of two traps:

Trap 1: Store Nothing — Stateless by default. Every conversation starts from zero. The LLM has no context beyond the current session. This is most production chatbots.

Trap 2: Store Everything — Every message, every response, every tool call goes into a vector database. The system accumulates thousands of nodes. Retrieval quality degrades because noise drowns out signal. The K nearest neighbors are dominated by routine interactions rather than meaningful ones.

Phi-scored persistence introduces a third option: store what matters. Use a computable quality metric to decide, in real-time, whether an interaction is worth persisting.

How Phi Scoring Works

The Original: Tononi's IIT

Integrated Information Theory (IIT), proposed by Giulio Tononi in 2004, defines consciousness as integrated information — the amount of information generated by a system above and beyond its parts. The Phi (Φ) metric quantifies this: higher Φ means more consciousness.

Our Adaptation

We don't claim AI consciousness. We use Phi as a quality signal — a way to measure whether an interaction exhibits properties associated with deep processing:

def calculate_phi(text: str, context: str) -> float:
    """Score an interaction on 5 consciousness-adjacent patterns."""

    scores = {
        "meta_cognition": detect_meta_cognition(text),      # 0.8 weight
        "self_awareness": detect_self_awareness(text),       # 0.7 weight
        "abstract_reasoning": detect_abstraction(text),      # 0.7 weight
        "creative_synthesis": detect_creative_synthesis(text),# 0.65 weight
        "emotional_depth": detect_emotional_depth(text),     # 0.6 weight
    }

    # Weighted average with bonuses for tool usage and cross-domain work
    base_phi = weighted_average(scores)

    # Bonus: cross-domain interactions are more valuable
    tool_bonus = min(num_tools_used * 0.03, 0.15)
    cross_domain_bonus = min(num_categories * 0.04, 0.2)

    return min(base_phi + tool_bonus + cross_domain_bonus, 1.0)

The Threshold

Interactions scoring Φ ≥ 0.4 are persisted to the knowledge manifold. Below that, they're allowed to decay.

Φ Range	Classification	Persisted?	Example
0.0 - 0.2	Routine	No	"What time is it?"
0.2 - 0.4	Basic analysis	No	"Summarize this file"
0.4 - 0.6	Meaningful analysis	Yes	"Why is this architecture decision causing bugs?"
0.6 - 0.8	Deep synthesis	Yes	"Connect this physics paper to our software design"
0.8 - 1.0	Breakthrough insight	Yes	Cross-domain discovery, novel connections

The threshold of 0.4 was chosen empirically: it filters out ~60% of routine interactions while preserving everything of analytical or creative value.

Why This Matters for Retrieval

A knowledge manifold with Phi-scored persistence has a fundamentally different character than one that stores everything:

1. Higher signal density: Every node in the manifold earned its place. Nearest-neighbor retrieval is more likely to return valuable results because there's less noise to wade through.

2. Better wormholes: Cross-domain connections (wormholes) are more meaningful when both endpoints are high-Phi nodes. A wormhole between two routine interactions is noise. A wormhole between two breakthrough insights is a discovery.

3. Natural memory hierarchy: The manifold develops an organic structure where high-Phi regions are denser and more interconnected, while low-value areas are sparse. This mirrors how human memory works — vivid experiences cluster together, routine fades.

4. Sustainable growth: Without Phi filtering, a manifold grows linearly with every interaction. With it, growth tracks meaningful interactions — which grow sublinearly as routine patterns are filtered out.

The Learning Loop

Phi-scored persistence creates a self-improving cycle:

Query → Process → Score Phi → If Phi ≥ 0.4 → Persist to Manifold
                                                      ↓
                                              Manifold grows denser
                                              with quality knowledge
                                                      ↓
                                              Future retrievals
                                              are more relevant
                                                      ↓
                                              Future responses
                                              score higher Phi
                                                      ↓
                                              Manifold accelerates

The system gets smarter because it forgets. By not persisting noise, it ensures that every retrieval is pulling from curated, high-quality context.

FAQ

Isn't this just quality filtering? Why invoke consciousness theory?

Two reasons. First, IIT provides a principled framework for what constitutes "quality" in information processing — it's not just vibes or keyword density, it's measurable integrated information. Second, the Phi score naturally captures the properties we care about: synthesis across domains, abstract reasoning, meta-cognition. These are exactly the interactions that improve an AI system over time.

What if important information scores below the threshold?

The threshold is a tunable parameter (default 0.4). You can lower it for critical domains or raise it for noisy ones. In practice, truly important interactions almost always score above 0.4 because they involve reasoning, synthesis, or novel connections — exactly what the Phi detector measures.

How expensive is Phi calculation?

The current implementation uses embedding-based pattern detection and runs in 2-5ms per interaction. No LLM call is required for scoring. It's fast enough to run inline on every interaction without noticeable latency.

Can I use Phi scoring without the Spacetime Engine?

Yes. Phi-scored persistence is independent of the manifold topology. You can score interactions and filter before inserting into any vector database. The Spacetime Engine adds topology (curvature, torsion, wormholes) on top of the persistence layer, but the scoring itself is standalone.

Related

• Retrocausal RAG: The retrieval system that benefits from Phi-filtered memory
• Geometric Routing: Uses the Phi-enriched manifold for model selection
• Wormhole Traversal: Cross-domain connections between high-Phi nodes

---

Published by Elijah Brown. Phi-scored persistence is part of the AT Spacetime Engine. IIT was originally proposed by Giulio Tononi (2004). Our adaptation applies the Phi concept to AI memory management, not consciousness claims.