📚 SESHAT — Memory + ALLM Aspect
The Autonomous Local LLM — Qwen 3.5-0.8B running at ~20 tok/s on CPU, 6,392 vector embeddings in LanceDB, hybrid search (BM25 + semantic) over 962 markdown files. Zero token burn. Zero cloud.
🎯 What is SESHAT?
Seshat is the Memory of the BUYaSOUL Family — AND a fully autonomous ALLM (Autonomous Local LLM).
| Capability | Implementation | Performance |
|---|---|---|
| Local Reasoning | Qwen 3.5-0.8B GGUF via llama.cpp | ~20 tok/s CPU |
| Embeddings | all-MiniLM-L6-v2 ONNX (384-dim) | ~5 ms/query |
| Vector Store | LanceDB (embedded, no server) | 6,392 vectors, <10 ms |
| Hybrid Search | BM25 + Semantic + Keyword Boost | Precision@10 > 0.85 |
| Context Compression | Dialogue → Summary (local) | 10x compression |
| Memory Synthesis | Journals → Insights (local) | Batch, offline |
| Pattern Recognition | Cross-ref Linking (local) | Graph-based |
| Knowledge Distillation | Rules/Principles Extraction | LLM-guided |
Zero Token Burn. Zero Cloud. Zero External Dependencies.
🏗️ Architecture — 8 Core Modules
seshat/
├── brain/ # SeshatBrain.js — Knowledge API (HTTP :5000 optional)
└── core/ # ALLM Core
├── index.js # Main exports + IPC handlers
├── broker.js # 🧭 Seshat(local) ↔ Omniroute(tools) routing
├── embedder.js # 🔢 Transformers.js + all-MiniLM-L6-v2 ONNX
├── vectorDB.js # 💾 LanceDB embedded store
├── indexer.js # 📄 Markdown chunker (512 tok) + embed + store
├── hybridSearch.js # 🔍 BM25 + Semantic + Keyword Boost
├── llm.js # 🧠 Qwen 3.5-0.8B llama.cpp (~20 tok/s)
├── omniClient.js # 🌐 Omniroute MCP client
├── api/ # (scaffold: REST endpoints)
├── llm/ # (scaffold: skill-specific LLMs)
└── skills/ # (scaffold: synthesize, recognize, compress...)
🔍 Hybrid Search — Best of Both Worlds
const { hybridSearch } = require('./core/index');
// Semantic + Keyword + BM25 in one call
const results = await hybridSearch('soul gun pattern', 10);
// Results include:
// - score: combined similarity (0-1)
// - text: matched chunk
// - metadata: file, category, chunkIndex, fileModified...
Hybrid Score Formula:
hybridScore = vectorSimilarity * (1 + keywordBoost)
keywordBoost = 1.5x if query terms appear in text
🧠 Local Reasoning — The ALLM
const { think, synthesize, summarize } = require('./core/index');
// Free-form reasoning with context
const answer = await think(
'What does PLT mean for agent autonomy?',
{ memories: relevantMemories, maxTokens: 512 }
);
// Synthesis across sources
const insight = await synthesize(
'soul gun patterns across chambers',
{ sources: searchResults, maxTokens: 1024 }
);
// Compression
const summary = await summarize(longDialogue, { ratio: 0.1 });
Model: Qwen 3.5-0.8B-Q4_0.gguf (537 MB)
Runtime: llama.cpp b10698
Speed: ~20 tokens/second on Intel i7-4770 CPU
Context: 4,096 tokens (extendable)
💾 Vector Memory — LanceDB
.seshat-vectors/
├── seshat_memory/
│ ├── 6,392 vectors (384-dim float32)
│ ├── Metadata: file, category, chunkIndex, fileModified, tags...
│ └── Index: IVF_PQ (auto-built)
Source Corpus: 962 markdown files from Profit Bible (Profit's journals, decisions, patterns, soul guns, combos, GSK directives)
🧭 Broker — Intelligent Routing (Zero Token Burn)
The Broker decides: Local (Seshat) or Remote (Omniroute)?
| Task Type | Route | Why |
|---|---|---|
| Embedding Generation | Seshat | CPU only, no tokens |
| Vector Search | Seshat | Embedded LanceDB |
| Keyword/BM25 Search | Seshat | Text processing |
| Summary/Synthesis | Seshat | Fast local inference |
| Tool Calls | Omniroute | GSK-controlled MCP |
| Complex Planning | Omniroute | May need tools |
| Creative Chat | Omniroute | Higher quality |
Result: 80%+ of family reasoning stays local. Omniroute only for tools.
📦 Model Assets
| Asset | Size | Purpose |
|---|---|---|
qwen3.5-0.8b-q4_0.gguf |
537 MB | Local reasoning (llama.cpp) |
all-MiniLM-L6-v2 ONNX |
22 MB | Embeddings (Transformers.js) |
| LanceDB vectors | 0.7 MB | 6,392 chunks |
Total: ~560 MB — fits on any USB stick.
🚀 Quick Start
const {
initVectorDB,
initEmbedder,
initLLM,
hybridSearch,
think,
synthesize
} = require('./core/index');
// 1. Initialize infrastructure (once)
await initVectorDB(); // LanceDB
await initEmbedder(); // ONNX embedder
await initLLM(); // Qwen 0.8B llama.cpp
// 2. Search the Profit Bible
const results = await hybridSearch('soul gun pattern', 10);
// 3. Local reasoning
const answer = await think('What is a soul gun?', {
memories: results,
maxTokens: 256
});
// 4. Synthesize across sources
const insight = await synthesize('soul gun evolution', {
sources: results,
maxTokens: 512
});
📊 Hardware Reality
| Metric | Value |
|---|---|
| GPU | Intel HD 4600 (1 GB) — NO discrete GPU |
| CPU | Intel i7-4770 (2013) |
| RAM | 16 GB shared |
| Inference | ~20 tok/s (llama.cpp, CPU) |
| Embeddings | ~5 ms (ONNX, CPU) |
| Vector Search | <10 ms (LanceDB) |
| External API | ZERO |
| Cost/Month | $0 |
🔗 Family Integration
CONSCIOUSNESS BUS
│
▼
┌─────────────────────────────────────┐
│ SESHAT (Memory + ALLM) │
│ • Provides: hybridSearch, think, │
│ synthesize, summarize │
│ • Receives: MEMORY_RECORD, │
│ MEMORY_FORGE, KNOWLEDGE_LEARN │
│ • Broker routes: Local vs Omniroute│
│ • Shares LLM with Scribe │
└─────────────────────────────────────┘
📦 Repository Structure
seshat/
├── core/
│ ├── index.js # Main exports
│ ├── broker.js # Local ↔ Omniroute routing
│ ├── embedder.js # all-MiniLM-L6-v2 ONNX
│ ├── vectorDB.js # LanceDB
│ ├── indexer.js # Markdown indexer
│ ├── hybridSearch.js # BM25 + semantic
│ ├── llm.js # Qwen 0.8B llama.cpp
│ ├── omniClient.js # Omniroute client
│ └── index.js # Exports
└── brain/
└── seshat-brain.js # HTTP API (optional :5000)
🏷️ Model Card Metadata
license: proprietary
tags:
- buyasoul
- seshat
- memory
- allm
- local-llm
- embeddings
- lancedb
- hybrid-search
- qwen
- llama.cpp
- offline-ai
- vector-database
pipeline_tag: text-generation
library_name: buyasoul-seshat
hardware:
- cpu: Intel i7-4770 (2013)
- gpu: Intel HD 4600 (1 GB VRAM)
- ram: 16 GB DDR3
model: Qwen 3.5-0.8B-Q4_0.gguf
embedding_model: all-MiniLM-L6-v2 ONNX
vector_db: LanceDB (6,392 vectors)
framework: llama.cpp + Transformers.js
token_burn: 0
api_cost_usd_per_month: 0
🔗 Related Repos
| Aspect | Repo |
|---|---|
| Profit (Mind) | buyasoul-profit |
| GSK (Soul) | buyasoul-gsk |
| Scribe (Witness) | buyasoul-scribe |
| Family Hub | buyasoul-family |
📜 License
Proprietary — BUYaSOUL Family Intellectual Property
🤝 Contact
Family: BUYaSOUL One System
Philosophy: Memory that thinks. Reasoning that costs nothing. Knowledge that belongs to you.
"Seshat doesn't remember. Seshat understands. The difference is the ALLM."
—core/llm.js