Memory Layer System: Multi-Level Memory for Agents

[sajosam]

Problem

Current memory in SpawnVerse is functional but flat:

• SQLite → shared state (short-term)
• Vector DB → semantic retrieval (optional)

Missing:

• structured memory hierarchy
• clear separation of memory types
• deeper reasoning continuity

---

Proposal

Introduce a multi-layer memory system, inspired by cognitive architectures:

1. Working Memory

   • current run state
   • agent outputs, messages
   • short-lived

2. Episodic Memory

   • past runs (fossils)
   • agent performance history
   • sequence of events

3. Semantic Memory

   • knowledge base (vector DB)
   • facts, documents, embeddings

---

Integration Idea

• Tag memory writes by type:

  • working
  • episodic
  • semantic

• Extend memory API:

write_memory(type="working", key="result", value=...)
read_memory(type="episodic", query="EV research")

• Auto-route:
  • results → episodic (fossils)
  • context → semantic (vector DB)
  • runtime → working (SQLite)

---

Benefits

• clearer reasoning flow
• better reuse of past runs
• improved RAG quality
• foundation for learning systems

[sajosam]

Right now we have memory.

This introduces structure — which is what enables real reasoning over time.