V3 Architecture

V3 Architecture

SuperLocalMemory V3 is a complete architectural reinvention — a mathematical retrieval engine built on information geometry, algebraic topology, and stochastic dynamics.

---

Overview

┌──────────────────────────────────────────────────────┐
│                 SuperLocalMemory V3                    │
│                                                        │
│  ┌──────────────────┐  ┌────────────────────────────┐ │
│  │  Product Shell     │  │  Mathematical Engine       │ │
│  │                    │  │                            │ │
│  │  CLI (15 commands) │  │  4-Channel Retrieval       │ │
│  │  MCP Server (24)   │  │  Fisher-Rao Similarity     │ │
│  │  Web Dashboard     │  │  Sheaf Consistency         │ │
│  │  17+ IDE Configs   │  │  Langevin Lifecycle        │ │
│  │  Learning (LightGBM│  │  11-Step Ingestion         │ │
│  │  Trust (Bayesian)  │  │  Scene + Bridge Discovery  │ │
│  │  Compliance (ABAC) │  │  Cross-Encoder Rerank      │ │
│  │  Profiles (16+)    │  │  3 Operating Modes         │ │
│  └──────────────────┘  └────────────────────────────┘ │
└──────────────────────────────────────────────────────┘

---

4-Channel Hybrid Retrieval

V3 retrieves memories through four parallel channels, each capturing different aspects of relevance:

Query
  │
  ├─ Strategy Classification (single-hop / multi-hop / temporal / open-domain)
  │
  ├─ 4 Parallel Channels:
  │  ├─ Semantic Channel (Fisher-Rao weighted embedding similarity)
  │  ├─ BM25 Channel (keyword matching, k1=1.2, b=0.75)
  │  ├─ Entity Graph Channel (spreading activation, 3 hops, decay 0.7)
  │  └─ Temporal Channel (3-date model: observation, referenced, interval)
  │
  ├─ Profile Lookup (direct SQL shortcut for entity queries)
  │
  ├─ Weighted RRF Fusion (k=60, channel weights vary by query type)
  │
  ├─ Scene Expansion (pull all facts from matched scenes)
  │
  ├─ Bridge Discovery (multi-hop only: Steiner tree + spreading activation)
  │
  ├─ Cross-Encoder Rerank (energy-weighted: α·sigmoid(CE) + (1-α)·RRF)
  │
  └─ Top-K Results with per-channel scores

Why Four Channels?

Channel	What It Catches	What It Misses
Semantic	Meaning similarity	Exact keywords, entity names
BM25	Exact terms, rare words	Paraphrases, synonyms
Entity Graph	Relational connections	Unconnected memories
Temporal	Time-relevant facts	Atemporal knowledge

No single channel handles all query types. The fusion combines their strengths.

---

Three Operating Modes

Mode	Description	LLM	EU AI Act
A: Local Guardian	Pure mathematical retrieval. Zero cloud calls.	None	Compliant
B: Smart Local	Mode A + local LLM (Ollama) for extraction.	Local	Compliant
C: Full Power	Mode B + cloud LLM + agentic retrieval.	Cloud	Partial

Mode A is architecturally unique: no other memory system achieves meaningful accuracy without LLM calls. The 4-channel retrieval + cross-encoder reranking provides high-quality results without generative AI.

---

11-Step Ingestion Pipeline

Every memory is processed through structured encoding before storage:

Step	What Happens
1	Entropy gating — information-theoretic filtering (low-entropy = skip)
2	Fact extraction — atomic, typed facts (episodic/semantic/opinion/temporal)
3	Entity resolution — canonical names with alias tracking
4	Temporal parsing — 3-date model (observation, referenced, interval)
5	Type routing — classify fact types for specialized handling
6	Emotional signal extraction — valence and arousal tagging
7	Knowledge graph construction — entities as nodes, relationships as edges
8	Consolidation — merge/update/supersede existing facts
9	Scene clustering — group facts by temporal-semantic coherence
10	Observation building — structured entity profiles
11	Foresight generation — anticipatory indexing for future queries

---

Database Schema

V3 uses a 17-table SQLite schema with FTS5 full-text search:

Core: profiles, memories, atomic_facts, atomic_facts_fts (FTS5) Entities: canonical_entities, entity_aliases, entity_profiles Graph: graph_edges, memory_scenes, temporal_events Quality: consolidation_log, trust_scores, provenance Learning: feedback_records, behavioral_patterns, action_outcomes Compliance: compliance_audit Infrastructure: bm25_tokens, config, schema_version

All tables are partitioned by profile_id for multi-context isolation (16+ profiles).

---

Code Structure

superlocalmemory/src/superlocalmemory/
├── core/           Engine, config, modes, profiles, embeddings
├── retrieval/      4-channel engine, semantic, BM25, entity, temporal, fusion, reranker
├── math/           Fisher-Rao metric, sheaf cohomology
├── dynamics/       Langevin lifecycle, Fisher-Langevin coupling
├── encoding/       11-step pipeline (entity resolver, fact extractor, scene builder...)
├── storage/        Database, schema, migrations, V2 migrator
├── compliance/     EU AI Act, GDPR, ABAC
├── learning/       Adaptive learning, behavioral tracking, outcomes
├── trust/          Trust scoring, provenance tracking, gates
├── llm/            LLM backbone (Ollama / Azure / OpenAI)
├── mcp/            MCP server (24 tools, 6 resources)
├── cli/            CLI with setup wizard (15 commands)
├── server/         Dashboard API + UI server
└── tests/          1400+ tests

---

Dashboard

slm dashboard    # Opens at http://localhost:8765

v3.4.4 "Neural Glass" Dashboard

Your browser does not support the video tag. Watch the demo.

21 tabs: Dashboard, Knowledge Graph, Memories, Recall Lab, Timeline, Clusters, Patterns, Learning, Behavioral, Live Events, Agents, Trust, Lifecycle, Compliance, Math Health, IDEs, Health Monitor, Ingestion, Entity Explorer, Mesh Peers, Settings.

Benchmarks

Evaluated on the LoCoMo benchmark — 10 multi-session conversations, 1,986 total questions.

Configuration	Aggregate	Multi-Hop	Open Domain
Mode A Retrieval (10 convs, 1,276 questions)	74.8%	70.3%	85.0%
Mode A Raw (zero-LLM)	60.4%	43.0%	72.0%
Mode C (conv-30, 81 questions)	87.7%	100.0%	86.0%

Ablation (conv-30, 81 questions)

Removed	Impact
Cross-encoder reranking	-30.7pp
Fisher-Rao metric	-10.8pp
All math layers	-7.6pp
BM25 channel	-6.5pp
Sheaf consistency	-1.7pp
Entity graph	-1.0pp

Mathematical layers contribute +12.7pp average across 6 conversations (n=832), with up to +19.9pp on the most challenging dialogues.

Full methodology and results in the V3 paper (Zenodo).

---

Part of Qualixar · Created by Varun Pratap Bhardwaj