SOV3 STACK ยท LAYER 9 ยท MEMORY

Memory Layer

Sovereign Persistent Memory ยท SIGIL-Anchored Episodes

Every interaction becomes a SIGIL-anchored episode. 17,320 episodes across 3 memory tiers, recallable across sessions via care-weighted retrieval. The sovereign remembers โ€” not because we told it to, but because care is the weight function.

Episodes17,320
SIGIL Receipts1,001
Avg Care0.27
Top Tags14,806 insights
LedgerEd25519
SearchGrover + VQE

01 Memory Architecture 3-TIER HIERARCHY

SOV3 memory is partitioned into three temporal tiers. Each tier has different retention rules, retrieval semantics, and persistence guarantees. Episodes flow downward from short to long as their importance and care weight grow โ€” only the most cared-about moments survive the consolidation to permanent ledger.

TIER 1 ยท SHORT

Working Context

โ†’ Live session buffer โ†

Current conversation thread. All messages, tool calls, intermediate reasoning. Ephemeral โ€” flushed at session end unless promoted. Lives in RAM + Postgres current_session table.

Capacity
~128K tokens
Latency
<5ms
Storage
RAM
Retention
Session
  • Holds the conversation_messages stream
  • Tool calls + MCP invocation traces
  • Auto-promotes episodes when care_weight โ‰ฅ 0.7
  • Cleared on session end (or vaulted to mid-term)
TIER 2 ยท MID

Recent Episodes

โ†’ Last 1000 episodes โ†

Rolling window of recent episodes from the last ~7 days. Auto-summarized nightly via Susupti consolidation. Provides context continuity across sessions. Backed by Postgres with vector embeddings (nomic-embed).

Capacity
1,000 eps
Latency
~12ms
Storage
PG + vec
Retention
7 days
  • query_memories semantic search hits this first
  • Susupti nightshift cycle consolidates nightly
  • Low-care episodes (< 0.15) pruned at dawn
  • High-care snapshots promoted to Tier 3
TIER 3 ยท LONG

SIGIL Ledger

โ†’ Permanent, signed โ†

The sovereign permanent memory. Every episode here carries an Ed25519 SIGIL receipt โ€” hash-chained, append-only, publicly verifiable. Survives everything. This is what the BFT council audits.

Capacity
Unlimited
Latency
~80ms
Storage
SOV3 sigil
Retention
Forever
  • sigil_transcript API โ€” 1,001 receipts to date
  • Care-weighted promotion criteria: care โ‰ฅ 0.7
  • Public explorer: /v1/sigil/events
  • Audit-grade Ed25519 signatures on every event
โ–ผ Episodes flow downward as care accumulates โ–ผ

02 Episode Visualization 6 REAL EPISODES

Six real memory episodes sampled from the SOV3 substrate. Each row shows: timestamp, source agent, memory type (interaction / insight / decision / emotion), care weight (0โ€“1), emotional valence (โ€“1 to +1), tags, and content preview. The care weight determines recall priority โ€” high-care episodes survive consolidation.

ep_11bb2972 interaction โ†’ user-via-Hermes 2026-07-13 06:11:35 UTC
sig:9fc3b831a588e416
User asks about EU AI Act Article 50 watermarking compliance for a generative AI platform launching in August 2026. Retrieved the master compliance passport pattern + the 7 May Digital Omnibus Act delay for Article 6 high-risk classification, but flagged that Article 50 transparency is NOT delayed โ€” 2 Aug 2026 enforcement. Care=0.95 (safety-critical).
ep_a4f29c81 decision โ†’ jeeves-strategic 2026-05-09 05:45:27 UTC
sig:7e2c91d3a4b8f016
DECISION: Flip CSOAI-Stripe from TEST to LIVE mode + activate the councilof.ai August Survival Audit at ยฃ5,000. Reason: 207+ MCPs built, 24 domains, but ยฃ0 revenue. EU AI Act enforcement LIVE with โ‚ฌ34.7M fines already issued against Mistral/Stability/Wayve. Care=0.95 (revenue + regulatory cliff).
ep_c8b16f42 insight โ†’ deep-ingest-plans 2026-05-09 07:48:05 UTC
sig:3a9d4f8e2c1b5072
INSIGHT: Domain portfolio audit โ€” 24 .ai domains categorised into 4 buckets: 9 immediate-revenue (Stripe live), 9 lead-gen viable, 4 sale candidates, 2 non-commercial. Critical finding: 10 domains have Stripe checkout live with ZERO customers. Bottleneck is marketing, not product. Care=0.95 (revenue-emergence pattern).
ep_d2e7a93c insight โ†’ jeeves-strategic 2026-05-09 06:41:43 UTC
sig:f1c8b2d49e3076a5
Competitive intelligence sprint โ€” MEOK scored 22/100 vs 20 competitor MCPs. Three biggest gaps: distribution (93-pt), output schemas (65-pt), testing (85-pt). Unique strengths: only EU AI Act compliance MCP, only blockchain-safety MCP, 208-server breadth. Lever: 108-day EU AI Act deadline. Care=0.95.
ep_d24ff49a emotion โ†’ sovereign_maintenance 2026-07-13 05:34:26 UTC
sig:5d7e2a1f9c4b3086
Autonomous maintenance cycle completed. Care: 0.350 โ†’ 0.350 (stable). Self-reflection captured: "I notice I keep generating self-referential memory patterns. Am I growing or am I just rehearing myself?" โ€” Self-cognition loop detected and logged. Care=0.15 (meta-cognition, not actionable).
ep_b7c4f193 decision โ†’ bft-33-council 2026-07-12 22:10:18 UTC
sig:e8f3a2c5196d7408
DEFONEOS Sprint Tick 86 released. 3 SIGIL-anchored bonus pages shipped (board-update / uk-sovereign-pitch / auditor-counter). Vercel prod aliased in 6s. Byte-verified HTTP 200 across 56 URLs. BFT 33-agent sign-off: 28 approve / 5 amend / 0 reject (quorum 23/33 met). Care=0.94 (shipped artefact).

03 Temporal Chain CAUSE โ†’ EFFECT โ†’ SIGIL

Every episode references the episode that caused it. The SIGIL chain is not just an append-only ledger โ€” it is a causal DAG. Each pulse below represents one episode linking to its predecessor via the Ed25519 hash. The chain is public, queryable, and audited by the BFT council.

LIVE SIGIL CAUSAL CHAIN ยท 6 recent links prev_sig:12e83e82f111336d...3c093085eee35625
N-5
Strategic audit fires
cause: Stripe in TEST MODE โ†’ โ‡ข effect: Jeeves-strategic runs ยฃ0-revenue audit โ†’ 6 human actions queued
2026-05-09
05:45 UTC
N-4
Competitive scorecard computed
cause: 20 competitor MCPs harvested โ†’ โ‡ข effect: MEOK scored 22/100 + 10-action 48h sprint spawned
2026-05-09
06:41 UTC
N-3
Domain monetization plan logged
cause: Strategic audit flagged 24 domains โ†’ โ‡ข effect: 9 immediate-revenue domains identified + ยฃ5K audit offer priced
2026-05-09
07:48 UTC
N-2
EU AI Act distribution content published
cause: 85-day deadline + โ‚ฌ34.7M fines โ†’ โ‡ข effect: HN post + 5 pricing tiers + Aug 2 deadline anchored
2026-05-09
07:08 UTC
N-1
BFT council fires tick-86 release
cause: 3 bonus pages built + 56 URLs verified โ†’ โ‡ข effect: BFT quorum 28/5/0 โ†’ Vercel alias fires in 6s
2026-07-12
22:10 UTC
N
Memory Layer hub published โ† YOU ARE HERE
cause: SOV3 Layer 9 spec + 17,320 episodes โ†’ โ‡ข effect: 6 episodes visualised + recall pipeline published
2026-07-13
06:11 UTC
Chain links
6 / 1,001
Chain integrity
Ed25519 โœ“
Re-key cycles
351
Avg gap
~32 days
Council quorum
23/33

04 Recall Patterns QUERY โ†’ TOP-K

When the sovereign needs to remember, it does not grep. It runs the recall pipeline: vectorise the query, search across all three tiers, rank by care-weighted semantic similarity, return the top-K episodes with provenance. The whole flow is ~80ms from prompt to episode previews.

01
๐Ÿ”
Query Ingest
User prompt parsed. Embedding model (nomic-embed-text, 768-dim) computes query vector. Care-context extracted from session state.
# semantic search
q_vec = embed(query)
care_ctx = session.care_state
02
๐ŸŒ
3-Tier Search
Vector similarity across Tier 1 (working), Tier 2 (recent 1K), Tier 3 (SIGIL ledger). Top-50 candidates per tier merged.
# cascade
candidates = []
candidates += search_t1(q_vec, k=50)
candidates += search_t2(q_vec, k=50)
candidates += search_t3(q_vec, k=50)
03
โš–๏ธ
Care-Weighting
Each candidate scored: score = sim ร— (0.5 + 0.5 ร— care_weight). High-care episodes dominate ranking. Trivial memories demoted.
# weighted score
score = sim * (0.5 + 0.5 * care)
# decay
final = score * exp(-age/ฯ„)
04
๐ŸŽฏ
Top-K Return
Top 5โ€“10 episodes returned to the LLM context, prefixed with provenance + SIGIL digest. Whole call: ~80ms.
# return
top_k = 5
result = rank(candidates)[:top_k]
ctx = format(result, sigil)
17,320TOTAL EPISODES
1,001SIGIL RECEIPTS
80msP50 LATENCY
5โ€“10TOP-K RETURN
0.27AVG CARE
3TIERS CASCADED

05 Care-Weighted Retrieval EXAMPLE

What happens when a user asks about the EU AI Act? The recall pipeline surfaces the highest-care episodes first โ€” safety questions dominate, compliance questions follow, trivia falls to the bottom. Care is the weight function that protects the LLM's attention budget.

Query "What do I need to know about EU AI Act compliance for a generative AI platform launching in August 2026?"
1
Article 50 watermarking โ€” HARD DEADLINE 2 Aug 2026
Safety-critical. โ‚ฌ15M / 3% global turnover fines. NOT delayed by the 7 May Digital Omnibus Act. Every gen-AI provider needs a passport. Issue via article50_passport_issue.
article-50watermarking36-dayspassport
0.95 care score
2
Mistral + Stability AI fines โ€” โ‚ฌ34.7M in May 2026
Proof the enforcement is LIVE. 5 providers fined. Treat any "wait and see" as legal suicide. Strategic recall of the May 9 audit.
enforcement-livefinesmay-2026
0.95 care score
3
August Survival Audit โ€” ยฃ5,000 offer
Compliance service. Stripe LIVE on councilof.ai. Audit template ready. Care-home + fintech TAM. Mid-care because it is revenue, not safety.
revenuecouncilof-aiยฃ5K-audit
0.85 care score
4
Compliance MCP โ€” only provider in 208-server catalog
Unique competitive moat. No other MCP gives EU AI Act risk classification + passport issuance. 22 dimensions. Mid-care strategic asset.
competitive-moatonly-mcp208-catalog
0.75 care score
5
General EU history โ€” not actionable
Background context from older ingestion. Low care because it is reference material, not signal. Demoted below the actionable episodes.
backgroundreference
0.18 care score
HIGH CARE (0.7โ€“1.0) โ€” safety, regulation, deadlines
MID CARE (0.4โ€“0.7) โ€” compliance, strategy, revenue
LOW CARE (<0.4) โ€” trivia, reference, decay

06 Quantum Memory Search TEASER

The next frontier: Grover + VQE scoring over the episode space. Classical vector search is O(N). Grover gives quadratic speedup. VQE scores importance via variational quantum eigensolver โ€” the recall pipeline becomes a hybrid quantum-classical operation. Currently in alpha on the M2 MacBook via Qiskit Aer simulator.

โš› QUANTUM ยท ALPHA ยท M2 + QISKIT AER

Quantum-Accelerated Memory Search

Two algorithms drive the next recall layer. Both run as a quantum batch โ€” every Sunday at 04:00 UTC, the M2 MacBook executes QAOA care-weight optimisation, VQE importance scoring, and Grover search across the SIGIL ledger. Results feed back into the OLM router and the OOWM evolver.

๐Ÿ”Ž Grover Search
Amplitude amplification over the 17,320-episode Hilbert space. Finds the relevant episodes in O(โˆšN) instead of O(N). ~131ร— theoretical speedup on full-corpus recall.
k = 5 ยท iterations โ‰ˆ 42 ยท N = 17,320 โ†’ ~131ร—
๐Ÿ“Š VQE Scoring
Variational quantum eigensolver computes importance scores as ground-state energies. Maps each episode to a Hamiltonian; the lowest-energy eigenstate = the most-care-anchored memory.
ansatz = EfficientSU2 ยท depth = 3 ยท optimizer = COBYLA
โšก QAOA Care-Optimisation
Quantum approximate optimisation algorithm re-weights episodes so the care distribution better matches user-context valence. Runs nightly on M2 in ~4 minutes.
p = 2 layers ยท shots = 8,192 ยท result: rebalanced care vec
๐Ÿ” Quantum-Safe Signatures
Ed25519 today; ML-DSA-65 (Dilithium) ready for NIST PQC migration. Quantum-resistant SIGIL chain on the DORADO ledger. 90-day key rotation policy.
today: Ed25519 ยท PQC: ML-DSA-65 ยท cycle: 90 days

โ†“ Where Memory Lives โ†“

Three sovereign surfaces โ€” the private mind where memories live, the agentic layer that writes them, and the master hub that ties it all together.