Sigma Runtime Standard - Public Specification Notice
This document is part of the Sigma Runtime Standard (SRS) public specification layer.Specification License: CC BY 4.0.
Implementation Safe Harbor: independent implementation permitted under public SRS/SRIP terms.
Machine-readable artifacts: Apache License 2.0 where explicitly marked.
Marks / Certification: governed by Sigma Marks and Certification Policy.
Proprietary Runtime Assets: not licensed by this SRIP.Independent implementations of public SRS/SRIP normative requirements are welcome under the public specification terms.
Product assets, protected Sigma marks, official certification, compatibility badges, CC BY-NC commercial use, and patent commitments use the relevant policy or explicit covenant. Independent implementation, attribution, or citation does not imply certification, endorsement, partnership, official compatibility, or permission to use Sigma marks as product identity.
| Field | Value |
|---|---|
| SRIP | SRIP-11 |
| Title | Compression and Memory Topology (CMT) |
| Version | 1.0 |
| Status | Active |
| Date | 2026-05-20 |
| Authors / Contributors | Sigma Stratum Research Group (SSRG) |
| Owning Layer | Memory / Compression / Topological Recall |
| Parent Specs | SRIP-09, SRIP-10 |
| Related Specs | None declared |
| Specification License | CC BY 4.0 |
| Implementation Safe Harbor | Independent implementation permitted under public SRS/SRIP terms |
| Machine-Readable Artifacts | Apache 2.0 where explicitly marked |
| Marks / Certification | Governed by Sigma Marks and Certification Policy |
| Proprietary Runtime Assets | Not licensed by this SRIP |
| Independent Implementation | Permitted under the public specification terms |
| Commercial Runtime Boundary | Relevant policy or explicit covenant for protected Sigma marks, official certification, managed deployment, white-label, resale, CC BY-NC commercial use, and patent commitments |
| Information Class | Open |
| Change Class | SRS-only |
| Normative Status | Defines a memory compression and topology contract. It does not mandate a specific vector store, graph store, compression scheduler, or recall product. |
| Conformance Level | Partial Conformance / Bounded Implementation |
| SRD Synchronization Action | Deferred review |
| Release Alignment Status | Active spec with historical applicability to SIGMA Runtime >= v0.5.3; no full runtime enablement claim is made by this document alone. |
Independent implementations of the public normative requirements in this SRIP are welcome under the applicable public specification terms.
No Sigma commercial runtime license is needed solely because an independent implementation follows those public normative requirements.
Product assets, protected Sigma marks, official certification, compatibility badges, CC BY-NC commercial use, and patent commitments use the relevant policy or explicit covenant. Independent implementation, attribution, or citation does not imply certification, endorsement, partnership, official compatibility, or permission to use Sigma marks as product identity.
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