Sigma Runtime Standard – License Notice
This document is part of the Sigma Runtime Standard (SRS).
It is licensed under Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0).The license for this specific document is authoritative.
See/legal/IP-Policyfor the full repository-wide licensing framework.
Version: 1.0
Status: Active
Applies to: SIGMA Runtime ≥ v0.5.3
Depends on: SRIP-09 (LTM), SRIP-10 (AEP)
Author: Sigma Stratum Research Group
Date: 2026-01-30
License: CC BY-NC 4.0 / Canon CIL Applicable
SRIP-11 defines the architecture of structural memory compression and semantic topology within Sigma Runtime.
Its goal is to reduce redundancy in long-term storage while preserving continuity of reasoning across extended cycles (≥ 500–5000).
Memory becomes topological: a self-organizing lattice of semantic nodes connected by relational weight rather than linear order.
| Term | Description |
|---|---|
| Rib Point | A periodic semantic condensation summarizing ≈ 10–50 cycles into a stable concept node. |
| Cluster | A set of rib points forming a thematic region (≈ 250 cycles). |
| Lattice | Graph of clusters connected by semantic edges (continuity vectors). |
| Continuity Vector | The direction of reasoning across cycles — derived from semantic gradient rather than text. |
| Density Coefficient | Local measure of information per token (symbolic density). Used for adaptive compression. |
| Phase Lineage | Inherited reasoning state (phase → phase) tracked through edges with type follows or transforms. |
| Anchor Facts Layer (AFL) | A priority memory layer for low-similarity but high-importance facts (names, identifiers, constraints). |
Every N cycles (default = 10 or 50):
Cycle → Embedding → Centroid → Rib Point → Cluster
Compression is not uniform:
records with high density and low redundancy are retained longer;
ephemeral ones are merged or pruned.
Retention weight = (density × coherence) / entropy.
Recall operates through graph propagation rather than flat similarity:
Memory is filtered by phase compatibility (forming, stable, reflection, fragmenting).
This prevents semantic cross-contamination between developmental stages of reasoning.
Phase compatibility matrix:
| Current Phase | Compatible Recall Phases |
|---|---|
forming |
forming, stable |
stable |
stable, reflection |
reflection |
reflection, stable, fragmenting |
fragmenting |
fragmenting, reflection |
Semantic similarity fails on certain anchor facts (e.g., names, identifiers, fixed constraints).
SRIP-11 defines a priority retrieval path that is domain-agnostic and always available in context.
AFL Responsibilities:
AFL Retrieval Order:
First N cycles are stored verbatim and always injected into context.
cmt:
anchor_buffer:
enabled: true
cycles: 8 # First N cycles always included as anchor context
The model extracts key facts dynamically — not limited to first N cycles.
Triggers (language-agnostic):
No hardcoded patterns. The model decides what to extract.
Fact Categories (universal):
identity: name, age, role, professioncontext: location, goals, time constraintsconstraints: limitations, allergies, restrictionsrelationships: family, colleagues, dependencieshistory: past events, backgroundcommitments: promises, deadlines, agreementspreferences: likes, dislikes, communication styleConfiguration:
cmt:
fact_extraction:
enabled: true
interval: 10 # Extract every N cycles
early_cycles: 3 # Always extract on first N cycles
min_confidence: 0.7 # Minimum confidence to store
max_facts: 30 # Maximum facts to store
# categories: [...] # Optional: custom categories (domain adapter)
# extraction_prompt: "..." # Optional: custom prompt
Safety: Only USER messages are used for extraction — assistant responses are context only.
This prevents the model from "inventing" facts from its own statements.
Context Injection Format:
[KNOWN FACTS — use these details in responses, especially summaries]
IDENTITY:
- name: Maria
- age: 52
- profession: accountant
CONSTRAINTS:
- allergy_penicillin: allergic to penicillin
RELATIONSHIPS:
- sister_condition: Hashimoto's thyroiditis
cmt:
domain_adapter: null # or one of the following:
# BASE
# - "conversational" # general assistant (recommended default)
# SPECIALIZED
# - "healthcare" # medical AI (names, allergies, medications)
# - "defense" # military/security (callsign, clearance, constraints)
# - "business" # enterprise (role, company, KPIs, stakeholders)
# - "education" # learning (level, goals, progress)
# - "legal" # legal (client, case, jurisdiction, deadlines)
# - "technical" # engineering (project, stack, architecture)
Adapters define domain-specific entity extraction patterns but do not alter core AFL logic.
When null, only anchor_buffer + fact_extraction operate universally.
Compression behavior can be customized per identity via the behavior: section in traits_*.yaml.
This section is NOT injected into the prompt — it controls internal system parameters only.
Configuration:
# traits_iaso.yaml (example)
behavior:
compression:
list_tolerance: 0.6 # 0.0-1.0: how much to allow lists (>= 0.8 = no penalty)
token_policy:
base_limit: 900 # Override default token limit
min_limit: 400 # Minimum allowed
max_limit: 1200 # Maximum allowed
Behavior:
list_tolerance >= 0.8 → skip list penalty entirelylist_tolerance < 0.8 → penalty scaled by (1 - list_tolerance)token_policy.base_limit → overrides system default max_completion_tokensUse Cases:
The AEP (Adaptive Entropy Protocol) monitors compression ratio and semantic loss per SRIP-10.
If semantic loss > threshold (default: 0.25), AEP triggers "phase regeneration":
→ expand compressed nodes temporarily for re-contextualization.
AEP integration points:
AEPController.monitor_compression() — evaluates CR and SL after each Rib Point creationAEPController.trigger_phase_regeneration() — queues Rib Point for expansion@dataclass
class RibPoint:
id: str # "RP-{start}-{end}"
cycle_range: Tuple[int, int] # (start_cycle, end_cycle) inclusive
vector: List[float] # Centroid embedding (384-dim for MiniLM)
summary: str # Compressed text representation
phase: str # Dominant ALICE phase during this range
density: float # Mean symbolic_density [0, 1]
entropy: float # 1 - mean(pairwise_cosine) [0, 1]
lineage: List[str] # Parent Rib Point IDs
created_at: float # Unix timestamp
metadata: Dict[str, Any] # Additional attributes
@property
def retention_weight(self) -> float:
"""Adaptive retention score: (density × coherence) / entropy"""
coherence = 1.0 - self.entropy
return (self.density * coherence) / max(0.01, self.entropy)
@dataclass
class Cluster:
id: str # "CL-{index}"
rib_points: List[str] # Member Rib Point IDs (5-10)
centroid: List[float] # Cluster centroid embedding
theme: str # Extracted theme label
phase_distribution: Dict[str, int] # Phase counts within cluster
cohesion: float # Mean pairwise similarity [0, 1]
created_at: float
Storage:
NetworkX directed graph with typed edges (follows, transforms, belongs_to, continuity)FAISS index for Rib Point centroids and Cluster centroids (negative cycle IDs)| Policy | Description |
|---|---|
| per_section | Compress each logical section (e.g., DN segment or topic). |
| per_cycle | Compress after N cycles of interaction. |
| hybrid | Alternate micro (10 cycles) and macro (50 cycles) compression. |
| structural | Triggered when density > threshold or semantic drift > 0.25. |
| Metric | Definition | Purpose |
|---|---|---|
| Compression Ratio (CR) | Tokens retained / Tokens input | Efficiency |
| Semantic Loss (SL) | 1 – cosine similarity(original, compressed) | Fidelity |
| Topology Cohesion (TC) | Average edge weight within cluster | Structural stability |
| Phase Continuity (PC) | Correlation of reasoning vectors across phases | Cognitive coherence |
| Anchor Recall Integrity (ARI) | Anchor facts retrieved / anchor facts declared | Reliability of AFL |
| Fact Extraction Rate (FER) | Facts extracted / extraction attempts | Quality of fact extraction |
| Fact Coverage (FC) | Categories with ≥1 fact / total categories | Breadth of fact capture |
End of SRIP-11 v1.0
Sigma Stratum Research Group – 2026