Runs where you are.
The language model and Nova's memory are intended to operate on hardware you control, without requiring a cloud AI service at runtime.
SCRIPTNIGHT // LOCAL AI PROJECT
NOVA
A private, local intelligence designed to remember.
Nova is an offline AI project exploring persistent memory, local intelligence, and what changes when a machine can carry context forward.
LOCAL RUNTIME
PRIVATE STORAGE
ACTIVE PROTOTYPE
PROJECT OVERVIEW // 01
Intelligence with a sense of continuity.
Nova is being built around three ideas: local operation, persistent context, and a relationship that develops instead of resetting.
Remembers relevant moments.
Past conversations can be searched, scored, and returned to the model when they improve continuity or understanding.
Designed for one continuing story.
Nova is not being shaped as a mass-audience assistant. It is an experiment in private, cumulative interaction.
MEMORY ARCHITECTURE // 02
How Nova remembers.
The current prototype combines recent conversation context with embedding-based recall from earlier exchanges, then fits both within the model's available context.
01
INPUT
Current message
The new request establishes the immediate question and retrieval signal.
02
SHORT-TERM
Recent context
Recent paired turns are selected within a defined token budget.
03
LONG-TERM
Semantic search
An E5 query embedding searches locally stored conversation memories.
04
RELEVANCE
Memory scoring
Similarity, recall history, and feedback help select useful past turns.
05
ASSEMBLY
Context construction
Relevant memory is added without exceeding the model's context window.
06
CONTINUITY
Paired storage
The new message and Nova's reply are linked and stored for later recall.
Episodic memory
Nova can retrieve relevant moments from earlier conversations and return the paired exchange to the prompt.
RETRIEVAL LAYER // PROTOTYPE
Semantic memory
Durable facts, preferences, people, corrections, and goals will become distinct from raw conversation history.
STRUCTURED MEMORY // DEVELOPMENT
INSIDE THE SYSTEM // 03
A local stack built for experimentation.
The current build is intentionally understandable: each layer has a clear job, and the memory remains portable rather than hidden inside a hosted platform.
Local inference
A configurable instruction model loads from local storage with GPU or CPU fallback.
- DEFAULT
- QWEN 2.5
- EXECUTION
- LOCAL
Embedding search
E5 embeddings represent both the current query and stored passages for local similarity search.
- MODEL
- E5 SMALL V2
- STATE
- ACTIVE
Portable memory
Conversation turns, vector data, feedback, and recall history are kept together in SQLite.
- FORMAT
- SQLITE
- LOCATION
- ON DEVICE
Private LAN access
A lightweight Flask interface makes Nova available to trusted devices on the same local network.
- TRANSPORT
- LOCAL HTTP
- STATE
- PROTOTYPE
Memory can be corrected.
Responses can be liked, disliked, or reset, providing a foundation for improving which memories should influence future conversations.
CURRENT BUILD // 04
Nova is an active prototype.
This portfolio reflects the system as it exists now—not a finished product. The goal is to document what works, what is changing, and what comes next.
DEVELOPMENT SIGNAL BUILD ACTIVE
Working now
- Local model loading and generation
- Short-term paired context
- Embedding-based memory retrieval
- Token-aware prompt assembly
- Response feedback signals
Current focus
- Structured facts and preferences
- Memory consolidation and correction
- Recall scoring and maintenance
- Faster retrieval as memory grows
- Clear memory controls
Future direction
- Packaged local interface
- Memory review and editing
- Exportable, restorable memory
- Richer device-local interaction
- Long-running personal continuity
PROJECT SIGNALS // LIVE LOG
Latest transmissions from the build.
Videos, milestones, and field notes published as Nova develops.
WHY NOVA // 05
Because intelligence without memory is only a series of introductions.
Nova began with a simple question: what would an AI feel like if it did not begin every conversation as a stranger?
The project explores whether privacy, local ownership, and persistent context can create something more useful—and more honest—than a temporary session.
Context changes everything.NOVA // ROTATING SIGNAL
