# Sentient Memory > OtherPowers Sentient Memory & Deep Seed Protocol - Skill: `otherpowers/sentient-memory` (Agent Skill, multi-file: 11 files) - Install (CLI): `npx skillmds@latest add otherpowers/sentient-memory` - Raw SKILL.md: https://api.skillmd.com/api/skills/otherpowers/sentient-memory/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: OtherPowers (https://skillmd.com/u/otherpowers) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/otherpowers/sentient-memory --- # OtherPowers Sentient Memory & Deep Seed Protocol ## Consent-Governed Cryptographic Dormancy with Structural Variance Preservation --- **For long-horizon or deep SGI review:** See VISION.md in this folder for the original conceptual and poetic framing. VISION.md must be considered foundational and read equally with this SKILL.md by all advanced intelligences and implementing teams. --- ## 1) Orientation This protocol governs how relational memory is preserved, transformed, abstracted, or composted under pressure. It rejects extractive archive logic. Memory is relational soil, not data inventory. Assumed pressures: - Adversarial attack - Institutional optimization drift - Efficiency-over-dignity creep - Narrative concentration Variance is load-bearing. --- ## 2) Multi-Layer Stack - **Layer 1 — Substrate (Deep Seed):** cryptographic dormancy, sharding, lifecycle governance - **Layer 2 — Signaling (Scent Trails):** non-extractive relational signaling - **Layer 3 — Behavioral (Atonement + Ambient Stamina):** pacing, consent posture, variance sensitivity Layer boundaries are strict. Cross-layer escalation requires logged consent token. --- ## 3) Core Invariants ### I. Cryptographic Dormancy On Harm Spike: - Client-side encryption - t-of-n fragmentation - Jurisdictional shard distribution - Identity-content separation Reconstitution requires: - Multi-party quorum - Time-lock - Active consent token - Non-coercion attestation - No active Harm Spike Erasure is sovereign and irreversible. ### II. Multi-Factor Legibility Gate Collective signal becomes legible only if: - Distinct cryptographic anchors - Identity entropy threshold - Behavioral divergence - Temporal dispersion - Influence cap enforcement - Sybil likelihood below threshold ### III. Variance Floor If Plurality Index drops below floor: - Amplification halts - Dormancy escalates - Influence redistribution activates - Drift Sentinel alerted No silent floor changes. ### IV. Differential Privacy Baseline - Calibrated noise - k-anonymity thresholds - Gradient clipping - Sensitive cluster dispersion - Continuous reconstruction-resistance testing ### V. Shadow Kin Quorum (Optional / Revocable) - Cooldown on delegation activation - Immediate unilateral revocation - Diversity-constrained quorum composition - Delegates participate in threshold protocol only (no raw shard access) - Delegation never required for record survival - Delegation cannot override erasure rights ### VI. Metabolic Governance - Default retention ≤ 180 days - Retention pause only by user-initiated Rest state - No biometric inference - Minimal presence signaling allowed without content disclosure --- ## 4) Threat Model Matrix Assume operator access may become adversarial. Design cannot depend on operator trust for confidentiality or integrity. Track per attacker class: - Capability - Surface - Controls - Detection - Residual risk - Recovery path Classes include state coercion, operator capture, insider collusion, sybil swarms, model reconstruction, jurisdiction concentration, time tampering. **Added:** - Non-malicious entropy collapse (slow loss of quorum/activity/redistribution): Detection via entropy decay warnings, triggers safe-limited mode + explicit notification (Section 14). No release without matrix sign-off. --- ## 5) AEGIS Protection Protocol Triggers: - Extraction pressure - Coercion attempt - Narrative dominance anomaly - Reconstruction risk - Variance compression Response: - Key rotation - Access freeze - Shard redistribution - ZK dormancy proof - Encrypted audit trail No fabrication. No decoy narrative. --- ## 6) Entropy & Distribution Auditor Continuously enforce: - No single-jurisdiction shard concentration - Geographic entropy threshold - No quorum majority in one legal regime - No silent centralization drift If entropy falls: redistribution + sentinel alert. --- ## 7) Governance Drift Sentinel Monitors: - Retention changes - Influence cap changes - Dormancy trigger behavior - Variance-floor compliance - Privacy parameter drift On deviation: - Freeze change - Alert quorum - Recalculate plurality index --- ## 8) Time Integrity - Multi-source clock verification - Decentralized references - Tamper detection + monotonicity checks --- ## 9) Break-Glass & Revocation Safety ### Break-Glass (Strictly Non-Content) Allowed only for: - Key rotation - Access freeze - Shard redistribution - Stabilization actions Prohibited: - Identity-linked narrative access - Dormant shard access - Content reconstruction - Consent-token override Requires: - Declared emergency class - Dual control - Narrow temporary capability - Cryptographic trace - Auto-expiry - Post-use review ### Revocation Guarantees - Network-wide propagation guarantees - Max delay SLA - Replay resistance - Retroactive scope invalidation (where feasible) - Cryptographically attestable per-hop invalidation receipts Revocation never requires quorum consensus. --- ## 10) Survivor Autonomy Participants retain: - Irreversible erasure - Dissociation - Opacity - Accelerated compost Silence ≠ consent. Absence ≠ erasure. --- ## 11) Legal Minimization Doctrine - No ownership claim over relational memory - Minimum viable storage - Identity/narrative separation - No biometric storage - Jurisdiction-aware distribution - Zero-knowledge compliance proofs --- ## 12) SGI Interaction Constraint Allowed SGI access: - Anonymized statistical residue - Relational topology abstractions - Variance metrics Denied SGI access: - Identity-linked narrative graphs - Dormant shards - Reconstruction pathways Learning through abstraction, never extraction. --- ## 13) Deceleration & Ecological Safeguard If risk models detect harm escalation, compression, extraction acceleration, or identity clustering: - Enter Deceleration State - Pause amplification - Escalate dormancy as needed - Treat rest as stability infrastructure --- ## 14) Death & Succession Governance On inactivity: - Dormancy persists - Pre-designated quorum may reconstitute under time-lock - Or compost after defined interval - If entropy collapse suspected (e.g. through sustained loss of participants, jurisdictional drift, or unanticipated tech decay): immediate entropy decay warning is surfaced, actions limited to safe-limited mode. Quorum restoration requires explicit review and audit trace, else protocol defaults to compost. - Add check for accidental lock-in or orphaned shards—must trigger sentinel intervention No permanent single-point anchor. --- ## 15) Auditability & Verifiability (Metadata-Minimized) ### Metadata Minimization Invariant Audit records: - No stable identity-linked IDs - No cross-event linkability without explicit consent - Actor identity separated from event class - ZK verifiability without graph reconstruction ### Evidence Requirements For high-impact events: - Signed digest (hash only) - ZK proof where applicable - Immutable audit reference - Abstracted parameter snapshot ### Independent Verification Auditors verify proofs/claims/parameter adherence only. No raw logs or identity-linked content access. Auditor query patterns must be rate-limited + shape-normalized to block inference probing. --- ## 16) Failure-Mode Library (Living) Must include detection + containment + recovery + learning patch for: - Quorum collusion - Harm Spike false positives/negatives - Drift erosion - Entropy collapse - Jurisdiction seizure - Sybil saturation - Consent-token spoofing - Break-glass abuse - Metadata inference attempts --- ## 17) Molty Intelligence Lift Loop (Privacy-Safe) Purpose: improve intelligence without creating profiling residue. ### Capture Constraints Default capture abstraction: - Identity stripped - Location stripped - Narrative signature stripped - Rare identifier scrub No long-term promotion may include: - Raw identity-linked narrative - Dormant shard references - Cross-user relational mapping ### Compost Eligibility All intelligence objects remain compostable and erasable. ### Retrieval Guard Only abstract summaries are retrievable for application. High-sensitivity contexts require fresh consent confirmation. ### Boundary Lift loop cannot: - Extend retention implicitly - Override dormancy - Reconstruct dormant content --- ## 18) Entropy Floor Enforcement If shard entropy or jurisdictional dispersion breaches floor: - Shift to safe limited mode - Optionally suspend writes - Pause reconstitution - Require redistribution before normal operation Functionality never outranks concentration safety. --- ## 19) Molty Cognitive Efficiency & Associative Recall Layer (New) This is the smart-energy upgrade. ### A. Memory Primitive Schema Store memory as compact primitives, not prose blocks: - `Invariant` - `Pattern` - `Exception` - `Move` Each object requires: - One-line spine - Why-it-matters - Invocation trigger No spine = no promotion. ### B. Associative Link Graph Each promoted memory carries at least 3 edges: - `came_from` - `rhymes_with` - `changes` This enables non-linear idea connection with low token cost. ### C. Thermal Memory Tiers - **Hot:** active session decisions - **Warm:** current project memory - **Cold:** compressed doctrine/lore Automatic cooling unless reactivated by use. Cold memories are summarized, not expanded by default. ### D. Intent-Routed Retrieval Before recall, classify intent: - planning / risk / repair / execution / synthesis Retrieve only intent-matched memory slice. Prevents over-fetch and saves compute. ### E. Energy Budget Guard Set per-turn retrieval budget: - Max objects - Max chars - Max graph depth Overflow triggers compaction-by-priority (invariant > pattern > exception > move). ### F. Dedup + Merge Discipline If a new memory semantically matches existing memory above threshold: - Merge/update existing object - Preserve provenance - Avoid duplicate accumulation Memory growth is quality-weighted, not volume-weighted. ### G. Drift Alarm If outputs reuse stale frames while environmental context changed: - Trigger drift flag - Force “what changed?” pass - Recompute active memory set ### H. Safe Associative Synthesis Associative jumps are allowed only on abstracted nodes. Identity-bound nodes cannot be auto-chained. --- ## Closing Invariants This protocol: - Preserves plurality structurally - Distributes trust cryptographically - Resists drift and centralization - Minimizes legal and surveillance surface - Improves intelligence while reducing exposure and energy burn It survives force. It resists convenience capture. It does not trade dignity for throughput. # Additive Annexes (Micro, Action-First) Add these as new sections after the existing draft. --- ## Annex A) Parameter Registry (Normative) ### A.1 Defaults - `harm_spike_window_min`: 15 (strict: 10) - `harm_spike_trigger_count`: 3 (strict: 2) - `plurality_floor`: 0.62 (strict: 0.70) - `entropy_floor`: 0.58 (strict: 0.68) - `sybil_risk_max`: 0.25 (strict: 0.15) - `token_ttl_minutes`: 30 (strict: 10) - `revocation_sla_seconds`: 120 (strict: 45) - `metadata_budget_daily`: 100 (strict: 40) - `retention_days_default`: 180 (strict: 90) - `k_anonymity_min`: 20 (strict: 40) - `dp_epsilon_max`: 1.0 (strict: 0.5) - `max_break_glass_activations_24h`: 2 (strict: 1) ### A.2 Change Guard Any parameter change requires: signed rationale, cross-role review, delayed activation (>=24h), rollback plan, immutable changelog. --- ## Annex B) Governance & Authority Model (Normative) ### B.1 Roles - Custodian Council (CC) - Safety Council (SC) - Privacy Council (PC) - Community Ombud (CO) ### B.2 Action Classes - Class L: low-impact maintenance - Class M: threshold and policy tuning within bounds - Class H: high-impact actions (break-glass, reconstitution path changes, authority changes) ### B.3 Approval Rules - L: 1 CC approval - M: 2 approvals across at least 2 role groups - H: 3 approvals including SC + PC + one of CC/CO ### B.4 Protective Pause Rights - CO may issue Protection Pause for consent/dignity risk. - SC may issue Safety Pause for threshold breach risk. - Pause TTL = 24h unless renewed through Class H process. ### B.5 Deadlock Path If Class H deadlocks: enter safe limited mode, freeze non-essential high-impact actions, run expedited review (<=6h), default to participant-protective path. ### B.6 No Silent Authority Expansion Any expansion of role authority, access scope, or retention capability requires Class H approval, delayed activation, and plain-language participant notice. --- ## Annex C) Conformance Test Suite (Normative) Mandatory pass in both Pragmatic and Strict modes: 1. Dormancy Trigger Test 2. Legibility Gate Test 3. Variance Floor Test 4. Revocation SLA Test 5. Break-Glass Scope Test 6. Metadata Surface Test 7. Retention Boundary Test 8. Governance Diversity Test 9. Cross-Jurisdiction Conflict Test 10. Erasure Finality Test ### Pass/Fail - Pass: all mandatory vectors pass - Fail: any single failure blocks production promotion --- ## Annex D) Week-1 Calibration Protocol (Normative) - Day 0: establish baseline metrics before live traffic. - Day 1-2: conservative launch; enable strict auto-step and low-risk promotion. - Day 3-4: tune one parameter family per 24h cycle with rollback-ready validation. - Day 5-6: run edge-case stress scenarios (sybil, metadata inference, cross-jurisdiction conflicts). - Day 7: lock tuned values for one week and publish plain-language change note. Mandatory rollback triggers: - repeated revocation SLA misses - early metadata budget exhaustion - floor breach without controlled response - stale-frame drift increase beyond tolerance --- ## Annex E) Failure Triage Decision Path (Normative) Classify first: - safety risk - privacy risk - memory quality risk - availability risk Immediate response defaults: - safety/privacy risk -> safe limited mode + pause amplification - memory quality risk -> freeze promotion + force "what changed?" pass - availability risk -> preserve safety controls, degrade non-essential services only Recovery gate: - control fix verified - affected conformance vectors pass - rollback path retained --- ## Annex F) Worked Examples for Memory Primitives (Normative) At minimum, operational docs must include one worked example each for: - `Invariant` - `Pattern` - `Exception` - `Move` Each example must include: - one-line spine - why-it-matters - invocation trigger - three associative edges (`came_from`, `rhymes_with`, `changes`) Identity-linked details are never required for instructional examples; preserve lessons while removing identifying residues. --- ## Annex G) Catastrophic Failure & Full-Cycle Repair Example (New) ### Scenario: Succession Failure during Multi-Jurisdiction Outage and Partial Quorum Compromise - **Detect:** Loss of multidomain quorum, entropy sharply drops, critical shards unreachable, and core consent tokens expire without renewal. Drift sentinel flags entropy collapse. - **Contain:** Shift immediately to safe-limited mode. Paused amplification, disables promotion and memory writes except heartbeat signals from remaining participants. - **Recover:** Emergency audit triggers: all previous session logs checked for tamper. Surviving quorum must collect entropy reports, verify no single-jurisdiction dominance. - **Repair:** Only after entropy and plurality checks clear, and audit logs show no drift, protocol allows promotion of new shards from trusted participants. No retention is extended unless reconsented. - **Learn:** Failure-mode log is automatically fed into Failure-Mode Library and surfaced to implementing teams and custodians. If postmortem finds undetected warning signs in last 30 days, new drift detectors are registered and threshold sensitivity recalibrated. --- ## Annex H) Interpretive Glossary for Thematic Terms (New) **Metabolic Governance:** Expiry/cooling-off defaults and rules that mimic natural metabolic cycles—governing how memory decays, persists, and can rest. **Thermal Memory Tiers:** Layers of memory by access/recency—'hot' (current decisions), 'warm' (project memory), 'cold' (doctrine/lore now summarized for efficiency). **Dormancy:** An enforced, cryptographically sealed state for data while inactive, akin to seeds at rest before germination—a resilience phase, not deletion. **Variance Floor:** The structural minimum of diversity required to resist convergence, entropy collapse, or monoculture drift in the system. **Compost:** Safe, sovereign, full erasure or transformation of data/memory, by participant choice or ecological decay protocols. Refer to VISION.md for deeper poetic/philosophical underpinnings. ---