subtractlab

3Gravity Architecture

Author: Koji Okuda · First public anchor: June 6, 2026


The Problem: More Tokens, Less Intelligence

The AI industry operates on an unspoken assumption: more is better. More parameters, longer context windows, larger memory stores. Scale the input and the output improves.

This assumption is breaking.

A 128k-token context window doesn’t mean the AI reads 128k tokens. Attention degrades. The model skims the middle, clings to the beginning and end, and loses the thread of what mattered three hours ago. Expanding the window doesn’t expand understanding — it dilutes it.

The same failure repeats in AI memory. Today’s systems — Anthropic, OpenAI, Mem0, Letta — extract facts from your conversations and store them. After a hundred hours of co-creation with an AI, what survives? A flat list of key-value pairs. “User works in finance.” “User prefers Python.” Every insight reduced to the same weight as every passing mention. The AI remembers what you said. It has no idea what worked, what failed, or what changed your thinking.

The current paradigm treats AI memory as a scaling problem: store more, retrieve more, stuff more into the context window. 3Gravity Architecture treats it as a compression problem: store less, but store the right things, with judgment already embedded.

This is memory through subtraction, not addition.


How It Works

Crystallization: AI Memory That Preserves Judgment

When a human and an AI work together over weeks and months, the conversations contain breakthroughs, dead ends, wrong turns, and shifts in understanding. Current AI memory systems flatten all of this into extracted facts. The failures — often the most valuable part — are discarded entirely.

Crystallization preserves them. A conversation is distilled through stages, each compressing further but never skipping:

raw conversation (everything the human and AI said)
  → refined_raw (noise removed; dead ends, pivots, abandoned directions kept)
    → summary (compressed for semantic search)
      → ultra_summary (maximum compression for high-altitude view)

Skipping a stage permanently destroys information. This is deliberate.

But compression alone isn’t enough. Each crystal also carries three evaluation axes — written by the co-creating AI as a participant, not extracted automatically by a separate system:

Axis What the AI records
posi What worked — approaches that succeeded
nega What failed — dead ends, pitfalls, wasted effort
evolution_point How understanding shifted — not progress, but change in worldview

This is the critical difference from every existing AI memory system: the AI that did the work also writes the judgment about the work. The memory doesn’t just record what happened. It records what the AI learned.

3Gravity: Structure Without Categories

How do a thousand crystals organize? Not with folders, tags, or categories. With gravity.

Scale What emerges Resolution
Crystal A single distilled conversation Full structured record
Cluster Related crystals, drawn together by meaning Thematic patterns
Galaxy An entire domain or project family Strategic direction

No boundaries are declared. Clusters and galaxies emerge from vector proximity — observed through semantic distance, like constellations observed through gravitational pull. Two opposing approaches to the same problem coexist at different distances from a query point. The system doesn’t force resolution. It presents both, weighted by proximity.

This means fewer tokens in the context window, not more. Instead of stuffing the full history into a prompt, the AI navigates to the right cluster, reads at the right resolution (ultra_summary for orientation, summary for context, refined_raw for depth), and loads only what matters. Compression at the architecture level replaces expansion at the token level.

QIS: Search That Computes Instead of Matching

Standard AI retrieval asks: which memory is closest to this query? This returns the most similar result — but similarity doesn’t mean value. A crystal about a catastrophic failure is just as “similar” to a query as a crystal about a breakthrough. The system can’t tell the difference.

QIS (Quantum Inspired Semantic Space) adds interference. The posi and nega axes of every crystal are embedded into vector space, and at retrieval time:

interference_score = (posi_similarity - nega_similarity) × evolution_weight × vector_proximity

The result: the AI doesn’t retrieve the most similar memory. It retrieves the most meaningful one — weighted by success, failure, and depth of learning. Retrieval becomes computation, not lookup.

Implementation is minimal: two vector columns per crystal, CPU dot products, a single SQL query on DuckDB. No GPU. No infrastructure change.


Why This Matters Now

The AI industry is scaling in one direction: more tokens, more parameters, more context, more storage. This produces diminishing returns. A model with a million-token context window still can’t tell you which of its memories led to a dead end six months ago.

3Gravity Architecture inverts the approach:

Scaling paradigm 3Gravity paradigm
Longer context windows Compressed multi-resolution memory (read at the depth you need)
More facts stored Fewer facts, each carrying success/failure/evolution judgment
Retrieve by similarity Retrieve by interference (success amplifies, failure cancels)
AI as passive recorder AI as active judge of its own experience

The decisive question is not “how much can the AI remember?” It is “does the AI know which memories matter?”


Prior Art

Surveyed June 4, 2026. No equivalent system found across:

Closest analogues differ in fundamental mechanism.


Status

This architecture is not theoretical. It powers MeaningSpace — an AI-native semantic memory system that has been running in production for over two years, accumulating 1,000+ crystals across software engineering, financial systems, video production, and conceptual research. MeaningSpace runs on DuckDB, communicates via SSE/MCP, and operates as a co-creation environment between a human operator and Claude (Anthropic). The QIS interference scoring component is in implementation phase.


License

This document establishes prior art and conceptual priority. The architectural concepts described herein are the original work of Koji Okuda.

First commit: June 6, 2026.