Project Planning

Memento Graph

MementoGraph is a personal intelligence substrate that continuously maps how a user's understanding evolves. It runs entirely locally, with Rust handling performance critical work and a local ~3B parameter language model (nomic-embed-text, 768-dimensional vectors) handling semantic extraction.

The system grows autonomously as content is ingested, discovering conceptual proximity, recurring themes, contradictions, and latent interest clusters without manual organization.

The interface exposes a dual layer cognitive architecture through two complementary visualization modes:

1. Graph View: the symbolic reasoning layer. Represents explicit links, causal structure, and contradictions as a semantic graph that grows organically with ingestion.

2. Semantic Field View: the associative memory layer. Uses UMAP dimensionality reduction to project high dimensional node embeddings into 2D/3D coordinates, rendering the resulting space as a landscape with density fields, cluster territories, drift paths, and boundary zones.

Architecture

The system is built in four layers:

1. Ingestion Engine (Rust/Tokio) — Parses content, chunks semantically, generates embeddings
2. Semantic Extraction (Rust + local LLM) — Produces structured Cognitive Objects via constrained decoding
3. Graph Engine (Rust + SurrealDB) — Manages the knowledge graph with autonomous relationship inference
4. Interface (Tauri + React/TypeScript) — WebGL graph visualization and exploration modes

Tech Stack

Layer	Technology
Systems / pipeline	Rust
Async runtime	Tokio
Local LLM inference	llama.cpp / ollama
Embedding model	nomic-embed-text (local, quantized)
Vector index	USearch (HNSW)
Graph database	SurrealDB (embedded)
Dimensionality reduction	UMAP (via linfa or umap-rs)
Desktop shell	Tauri
Frontend	React + TypeScript
Graph rendering	Sigma.js / Cosmograph (WebGL)

Project Structure

See docs/source-code-structure.md for the full layout.

crates/
├── memento-core/           # Shared types, traits, errors
├── memento-ingestion/      # Content parsing, chunking, embedding
├── memento-extraction/     # LLM backends, constrained decoding
├── memento-graph-engine/   # SurrealDB, inference, temporal weighting
└── memento-app/            # Tauri binary + React frontend

Development

Build

cargo build
cd crates/memento-app/frontend && npm install && npm run build

Run

cargo run -p memento-app

Documentation

• Requirements
• Technical Design
• Implementation Tasks