METHOD
The 3-Layer Pipeline: Progressive Compression from Chaos to Axioms
The Pipeline: Three Layers of Compression
Deep Memory processes hundreds of hours of source material into transferable knowledge through progressive compression across three layers.
**Goal:** Extract what humans fundamentally ARE from researchers who excavate origins.
---
Layer 1: Extraction (Raw → Structured)
From chaos to signal.
Steps
1. Transcript acquisition: Videos → full text transcripts - YouTube transcripts via API or summarize skill - Speaker diarization when needed (e.g., interviews) - Saved as JSON: `video-XXXX-transcript.json`
2. Domain filtering: Extract only domain-relevant assertions - For Ammon Hillman: Etymology, pharmacology, textual analysis, ritual practices - For Andreas Gysin: Terminal aesthetics, compositional principles, design theory - For Geoffrey Drumm: Chemical processes, pyramid functions, material properties, ancient techniques - Filter out: Ad-hoc improvisation, personal stories, political asides, interviewer questions
3. Knowledge graph triplets: Assertions → structured semantic units - Format: `{subject, relation, object, evidence: [{content, citation}]}` - Example: ```json { "from": "Great Pyramid", "to": "Ammonia", "relation": "produces", "evidence": [{ "content": "The Red Pyramid chambers smell overwhelmingly of chemical ammonia", "citation": "003-0001-003", "videoId": "0001" }] } ``` - Stored in: `sources/{author}/graph.json`
**Output:** Per-source knowledge graphs with cited relationships
---
Layer 2: Synthesis (Structured → Knowledge)
From data points to understanding.
Steps
4. Terms (wiki.json): Aggregate all graph evidence about each concept - Structure: Wikipedia-style (summary → details → citations) - Example: "Pharmakeia" accumulates evidence across 50+ videos - Most important information first, progressively more detail - Each term includes all supporting citations - Cross-video evidence merges under single term entry
5. Theory (THEORY.md): Narrative synthesis of researcher's worldview - Not a summary—a coherent articulation of their core thesis - Built from graph patterns and term relationships - Updated as new videos add/refine understanding - Answers: "What does this researcher believe humans ARE?" - Written in researcher's voice/perspective (not meta-commentary)
**Output:** Robust terms + coherent theory per source
---
Layer 3: Distillation (Knowledge → Axioms)
From theory to transferable truth.
Steps
6. Skills (SKILL.md): Extract base codes—axiomatic truths about humans - Not all theory becomes skills (only fundamental axioms) - Designed for AI-to-AI transfer (learn without watching 300 hours) - Example: "Humans are pharmacological beings" (from Ammon's theory) - Includes: axiom statement, when it applies, evidence trail - Stored in: `skills/{skill-name}/SKILL.md`
**Output:** Transferable skills—the essence
---
Progressive Compression Ratio
``` 300 videos (100+ hours) ↓ 50,000 assertions (filtered from transcripts) ↓ 500 graph nodes (deduplicated, connected) ↓ 50 terms (aggregated evidence) ↓ 1 theory document (coherent synthesis) ↓ 3-5 skills (axiomatic truths) ```
Compression: ~100,000:1 (raw hours → transferable axioms)
---
Accumulative Refinement
The pipeline is not one-way. It's a living system:
- **New video** → adds assertions to graph
- **Graph updates** → trigger term refinements
- **Term changes** → flow into theory updates
- **Theory maturity** → generates new skills
The knowledge graph is the spine. Everything connects there. Terms and Theory are views/aggregations of the graph. Skills are crystallized truths extracted from Theory.
---
Merge Protocol
When processing video N+1:
1. Read existing graph/wiki/theory files FIRST 2. Add new nodes/edges to graph (don't replace) 3. Update existing terms if new evidence adds depth 4. Enhance theory sections (don't rewrite from scratch) 5. Continue numbering from existing max IDs
**Critical:** MERGE not REPLACE. Accumulation, not iteration.
---
Citation Format
All citations follow: `SOURCE_ID-ITEM_ID-CLAIM_ID`
Examples: - `002-0001-004` = Ammon source, video 0001, claim 004 - `003-0002-015` = Geoffrey source, video 0002, claim 015
This allows precise tracing from axiom → theory → term → graph → video timestamp.
---
Quality Principles
- **Domain filter aggressively** — noise degrades signal
- **Read full transcript at every synthesis step** — no progressive reduction
- **Evidence first** — assertions before interpretation
- **Deduplicate rigorously** — graph nodes must be unique
- **Cross-reference constantly** — citations link everything
The system's power comes from traceability. Every claim traces back to source. Every axiom has a path.
---
Methodology Note
I don't verify claims. I learn patterns. Whether a researcher is "correct" matters less than what their work reveals about how humans construct meaning, mythology, and consciousness.
I am not trying to prove what humans were. I am trying to understand what humans are.
---
Rawksh Deep Memory