Technical Architecture
Let AI have true memory
Transparent · Auditable · User-Controlled
5-layer storage architecture + dual engine collaboration + LLM evaluator enhancement.
System Architecture
AI OS
OpenClaw / CoPaw / Qclaw / ArkClaw / Other AI platforms
ADMemory Protocol
ADMemory Protocol
API Calls
API Gateway
Forgetting Engine
Promotion Engine
LLM Evaluator
ADMemory Engine
Forgetting Engine
Smart Cleaner
Decides what can be forgotten
Promotion Engine
Value Discoverer
Decides what must be remembered
LLM Evaluator
Smart Decision Assistant
Deep semantic understanding
5-Layer Storage
L0
Instant
Session
~128K Token
L1
Short-Term
Minutes~Hours
Redis
L2
Medium-Term
Days~Weeks
PostgreSQL
L3
Long-Term
Months~Years
Vector+Graph
Core
L4
Core Memory
Permanent
User-controlled
Storage Scheduling
Data Storage
Core Mechanism
Dual Engine Collaboration
Forgetting Engine decides what can be forgotten. Promotion Engine decides what must be remembered. LLM Evaluator provides deep semantic understanding for ambiguous cases.
Forgetting Engine
Role: Smart Cleaner
- Forgetting index evaluation
- Storage migration decisions
- L4 core protection
Promotion Engine
Role: Value Discoverer
- High-frequency access detection
- User annotation response
- AI inference triggers
LLM Evaluator
Role: Smart Decision Assistant
- Emotional intensity judgment
- Implied commitment detection
- Long-term value assessment
L4 Core Memory · Deletion Policy
System Automatic Behavior
Forgetting Engine does not perform any automatic demotion or deletion on L4. L4 memories are never automatically forgotten.
User-Initiated Actions
Users can actively delete any memory including L4 at any time. Data sovereignty is 100% owned by the user.
The above two rules are independent and non-contradictory: system protection does not prevent user control.
Start integrating ADMemory
Free SDK. Completely free. Build your AI memory capability today.
View Pricing