Memory System¶

Neuralgentics does not use a simple "vector search" for context. It implements a Trust-Weighted Semantic Memory system that separates raw data from verified patterns.

📈 The Trust Engine¶

Unlike standard RAG, which treats every retrieved chunk as equally valid, Neuralgentics tracks the reliability of memories.

   RAW MEMORY
       │
       ▼
 ╔══════════════════╗
 ║  INITIAL TRUST   ║ ◄── All new memories start at 0.5
 ╚══════════════════╝
       │
       ▼
 ╔══════════════════╗       ▲ [ +0.05 ]  Agent successfully used memory
 ║  SIGNAL ENGINE   ║ ◄─────┤ [ +0.10 ]  User explicitly confirmed
 ╚══════════════════╝       ▼ [ -0.05 ]  Agent ignored the memory
                            ▼ [ -0.10 ]  User corrected the memory
       │
       ▼
 ╔══════════════════╗
 ║   DECAY ENGINE    ║ ◄── Trust slowly fades if never used
 ╚══════════════════╝
       │
       ▼
 [ ARCHIVE / PURGE ]  ◄── Trust < 0.3

Diagram 6 — Trust Scoring Pipeline. Every piece of information is a "hypothesis" until proven useful. Trust signals enable the system to remember how to solve a problem (high trust) while ignoring failed attempts from previous sessions.

층 Tiered Memory Loading¶

To prevent context window saturation, Neuralgentics loads memory in three distinct tiers:

    SESSION START
          │
          ▼
 ╔══════════════════════════════════════════╗
 ║ TIER 0: GLOBAL SUMMARY (~100 tokens)     ║ ◄── high-trust project context
 ╚══════════════════════════════════════════╝
          │
          ▼
 ╔══════════════════════════════════════════╗
 ║ TIER 1: KEY DECISIONS (~2K tokens)       ║ ◄── trust ≥ 0.8; laws/rules
 ╚══════════════════════════════════════════╝
          │
          ▼
 ╔══════════════════════════════════════════╗
 ║ TIER 2: FULL SEMANTIC SEARCH             ║ ◄── L2 full-vector retrieval
 ╚══════════════════════════════════════════╝

Diagram 7 — Tiered Memory Loading. This pyramid ensures the agent always knows the overall goal (L0) and the golden rules (L1) before diving into the specific technical details of a file (L2).

🕸️ Knowledge Graph (KG)¶

Beyond vectors, Neuralgentics tracks relationships between entities (Projects, Files, Agents, Skills).

   ┌──────────┐                      ┌──────────┐
   │ PROJECT  │ ────── RELATED ──────► │   AGENT  │
   └──────────┘                      └──────────┘
         │                                  │
         │ SUPERSEDES                       │ DERIVED_FROM
         ▼                                  ▼
   ┌──────────┐                      ┌──────────┐
   │ MEMORY A │ ◄── CONTRADICTS ──── │  MEMORY B │
   └──────────┘                      └──────────┘

Diagram 12 — Knowledge Graph Entity Model. The KG allows the system to handle contradictions. If Memory B is marked as SUPERSEDES Memory A, the orchestrator will ignore A even if its vector similarity is higher.

Relationship Types¶

Type	Description
`SUPERSEDES`	Replace an old decision with a new one.
`PARTIAL_UPDATE`	Add context to an existing memory.
`RELATED_TO`	Semantic connection without hierarchy.
`CONTRADICTS`	Explicit conflict (triggering a dialectic resolution).
`DERIVED_FROM`	A memory based on another (e.g., summary $\rightarrow$ detail).