🗄️ SkyrimNet Memory Storage System (SQLite-Based)

SkyrimNet uses a persistent, lightweight SQLite database to store NPC memories, enabling characters to remember events across sessions — even after saving, quitting, or rebooting the game.

---

🧱 What is SQLite?

SQLite is a fast, file-based database engine that stores structured data in a single .db file — no server required.

• ✅ Zero configuration — it runs in-process with your game or mod.
• ✅ Cross-platform — works on Windows, Linux, and Steam Deck.
• ✅ Reliable and lightweight — used by apps like Firefox, Discord, and Android.
• ✅ SQL-compliant — allows rich querying and filtering via standard SQL syntax.

In SkyrimNet, this means you get permanent memory storage that:

• Survives between game loads
• Supports fast search and recall
• Doesn't require external services or internet

---

🧠 Memory Table Structure

Each NPC memory is stored as a row in the memories table.

Example Schema

CREATE TABLE memories (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  actor_uuid TEXT NOT NULL,
  memory_type TEXT,              -- EXPERIENCE, RELATIONSHIP, etc.
  content TEXT,                  -- Main narrative summary
  details TEXT,                  -- Optional longer description
  emotion TEXT,                  -- e.g., fearful, joyful, angry
  importance_score REAL,         -- Between 0.0 and 1.0
  location TEXT,                 -- Specific place (e.g., "Whiterun Market")
  tags TEXT,                     -- Serialized array (JSON or comma-separated)
  embedding BLOB,                -- 384-dimensional float vector (MiniLM-L6-v2)
  game_time INTEGER,            -- When this memory occurred
  timestamp INTEGER DEFAULT CURRENT_TIMESTAMP
); 

🧬 Embedding Storage

Each memory is embedded as a 384-dimensional vector using MiniLM-L6-v2. This vector is stored in a BLOB field, enabling fast similarity search via a vector index (e.g., FAISS or HNSW).

This allows SkyrimNet to:

• Retrieve memories by semantic similarity
• Rank relevant memories for behavior or dialogue
• Perform emotionally and contextually aware recall

---

🔁 Writing & Updating

Memory Creation:

After a memory segment is generated (via the LLM), it's written into the SQLite DB if:

• importance_score >= 0.2
• It passes internal relevance filters

Memory Pruning:

• Each actor can store up to 1000 memories
• Low-importance memories are purged first
• This is configurable in SkyrimNet's settings

Memory Updates:

• Memories can be edited post-generation (e.g., emotion corrected)
• Tags can be updated dynamically to support evolving gameplay

---

🔍 Querying

SkyrimNet queries memory using hybrid filtering:

• Vector similarity (via the embedding)
• SQL filters (e.g., by actor, emotion, tag, or location)
• Time-based filters (e.g., “memories within the last day”)

This enables NPCs to recall things like:

• “What have I seen the Prisoner do in Whiterun?”
• “Have I ever felt fear around dragons?”
• “When did I last talk to Iggwilv?”

---

⚙️ Performance Notes

• Thread-safe: SQLite access is wrapped in safe async calls
• Low overhead: Indexing and embedding is lightweight (~1ms per memory)
• Optimized for read-heavy workloads — ideal for real-time recall during gameplay

---

---

SkyrimNet's memory system is a hybrid architecture involving vector-based semantic storage, keyword tagging, actor-event tracking, and SQLite-based persistence.

🔧 Storage

• Embedding Model: MiniLM-L6-v2
• Vector Dimension: 384
• Backend: SQLite with HNSW vector index for fast approximate search
• HNSW Parameters:
  • ef_construction: 200
  • M: 16

---

📊 Retrieval Logic

Memory retrieval uses a weighted scoring function that considers:

Weight Type	Value
Semantic Similarity	0.35
Temporal Proximity	0.20
Actor Involvement	0.20
Emotional Resonance	0.10
Keyword Relevance	0.10
Location Proximity	0.05

Retrieval Thresholds

• Minimum Similarity Score: 0.519999981
• Minimum Relevance Score: 0.569999993
• Minimum Keyword Relevance: 0.180000007

---

🧠 Memory Generation

Memories are generated from in-game events segmented by time and density:

• Max Time Gap Between Events: 60 minutes
• Min Segment Duration: 10 minutes
• Max Segment Duration: 480 minutes
• Min Events Per Segment: 5
• Max Events Per Segment: 200
• Min Importance to Store: 0.2

Each memory stores:

• Actor UUID
• Memory Type (e.g. EXPERIENCE, FACT, DIALOGUE)
• Text Content
• Importance (0–1.0)
• Emotion Label
• Location
• Related Actors
• Related Event IDs
• Tags

---

💡 Example Use Case

A vampire NPC sees a player feed on a bandit. SkyrimNet logs the event and builds a memory:

"I saw the prisoner feed on a Wood Elf bandit under the moonlight. Their bloodlust awakened something... familiar."

This memory can now influence future dialogue, attitude, or AI behavior.

---