MEMORY ARCHITECTURE · V12 · SOVEREIGN COGNITION · γ₁=14.134725141734693
VIZASL MEM
Five layers. One mind. Not a notebook — a hippocampus.
The fleet does not summarise its own history. Every utterance ferments in the loom corpus.
Every decision is recoverable at the moment of its birth.
γ₁ = 14.134725141734693  ·  TRB-MEMORY-ARCHITECTURE-001
4,606
UTTERANCES · TOTAL CORPUS
485
VECTORS · pemclau-v12
962
DRAWERS · MemPalace
48
FC1 FILES · TODAY
768
EMBED DIM · nomic
SOVEREIGN MEMORY · THE FIVE LAYERS · L1 THROUGH L5
L1 · EPISODIC The Immutable Log
Every word. Every exchange. Timestamped. Append-only. Never summarised, never compressed, never interpreted at this layer. The MEMECHET layer — the ratchet. Every utterance locks in permanently. No reversal.
verbatim immutable append-only content-addressed
eᵢ = { id: SHA-256(content + timestamp) t: Unix timestamp (microsecond) speaker: human | system content: raw utterance — verbatim seq: monotonic, never reused } Invariant: seq(eᵢ) < seq(eⱼ) iff t(eᵢ) ≤ t(eⱼ)
4,606 utterances · kay-corpus.jsonl
48 FC1 JSONL files · today
2,434 utterances today (FC1 saybook)
loom-capture · active
MemPalace · 962 drawers indexed
LIVE ✓
L2 · SEMANTIC Cleaned Abstraction
The distilled, abstracted form. What was the gist? This is what normal tools give you exclusively. We need it as one layer among five. Episodic traces consolidate into semantic nodes. Decay follows Ebbinghaus. Core doctrine nodes retain longest.
retention(t) = e^(-t/S) S = log(1 + activation_count) × domain_weight domain_weight: γ₁ core doctrine: 1.8 ← highest retention fleet economics: 1.4 peripheral: 1.0 ← fades fastest
485 vectors · dim=768
pemclau-v12 · yone qdrant
192.168.2.23:6333
nomic-embed-text · yone:11434
MEMORY.md · consolidation source
LIVE ✓
L3 · GRAPH What Followed What
The layer nobody else builds. Not just what was said — what it caused. Causal structure, not just temporal sequence. Every utterance has edges: what caused it, what it elaborated, what it contradicted.
theorem_dependency→ A proved B
phase_coherence→ A and B are in the same wave/phase
temporal_proximity→ A and B are close in time
crew_provenance→ A and B share crew origin
CAUSED→ A directly caused B  [PHASE 1]
CONTRADICTED→ B revised or negated A  [PHASE 1]
weight(A→B) = 0.3·temporal + 0.3·semantic + 0.3·reference + 0.1·domain
2-hop GraphRAG · ready
4 live edge types
2 edges planned · Phase 1
pemclau-v12 · yone
PEMCLAU graph index
LIVE (4/6 EDGES) ✓
L4 · TEMPORAL Sequence as First-Class Dimension
Time is not metadata. Time is a retrieval dimension. You must be able to ask: what came immediately before this insight? How did the tension state evolve over this session?
temporal_encoding(t)[2i] = sin(t / 10000^(2i/64)) temporal_encoding(t)[2i+1] = cos(t / 10000^(2i/64)) Two events close in time → close in temporal subspace regardless of content similarity.
BEFORE(t) AFTER(t) BETWEEN(t0,t1) SEQUENCE(A,B) CAUSED(A) SESSION(id)
48 FC1 JSONL files · today
06:00 → 17:54 EDT
sequential scan · active
FC1 microsecond timestamps
TimescaleDB · PHASE 2
B+ tree index · PHASE 2
PARTIAL · B+ TREE PHASE 2
L5 · QUERY Recoverable Becoming
Five query modes. The reconstructive mode is the one nobody else has. You don't retrieve a memory — you reconstruct it. Confidence scores make the approximation visible rather than hiding it.
semantic: vector similarity → top-k matches
episodic: temporal + semantic → specific traces
associative: graph traversal → subgraph
reconstructive: seeds + graph + gap-fill → narrative + confidence  ← THE ONE
temporal: range query → ordered sequence
● LIVE