ADMemory
Infrastructure for the AI OS Era
5-layer memory architecture + dual engines + LLM evaluator. Free SDK. Open protocol.
Memory lifecycle from instant to permanent
| Layer | Name | Lifetime | Description |
|---|---|---|---|
| L0 | Instant Memory | Session | Real-time context window. The entry point for all memory. |
| L1 | Short-Term Memory | Minutes ~ Hours | Working memory. Summary of recent interactions. |
| L2 | Medium-Term Memory | Days ~ Weeks | Important recent facts. Promoted from L1 or demoted from L3. |
| L3 | Long-Term Memory | Months ~ Years | Deep personal knowledge. Cross-session accumulation. |
| L4 | Core Memory | Permanent (Deletable) | Essential identity info. System-protected. User has full control. |
The intelligence behind the memory
Forgetting Engine
Determines what can be safely forgotten. Balances storage efficiency with memory retention. Never loses important data.
Promotion Engine
Identifies information worth promoting to deeper memory layers. Based on frequency, importance, and recency signals.
LLM Evaluator
For ambiguous cases, uses LLM's semantic understanding to assess memory importance. Handles nuance that rules-based systems miss.
Multiple ways to integrate
RESTful API
Simple HTTP interface. Best for web integrations.
gRPC
High-performance binary protocol. Best for real-time apps.
WebSocket
Bidirectional streaming. Best for live AI interactions.
Why ADMemory Stands Out
Compared with mainstream memory frameworks, ADMemory leads in architecture depth, mechanism completeness, and protocol standardization
| Feature | ADMemory | Mem0 | Zep | Letta | Supermemory | SuperLocalMemory |
|---|---|---|---|---|---|---|
| Memory Hierarchy | 5 Layers L0~L4 | 1 Layer | 2 Layers | 2 Layers | 1 Layer | 3 Layers |
| Forgetting Mechanism | Smart + LLM | Vector TTL | TTL Expiry | LLM Self | RRF Fusion | TTL Expiry |
| Promotion Engine | ✓ | — | — | — | — | — |
| LLM Evaluator | ✓ | — | — | — | — | — |
| Graph Database | L3 Planned | ✓ | ✓ | — | — | ✓ |
| Protocol Standard | Open Protocol | — | — | — | — | — |
| Benchmark | TBD | ~66% | ~85% | ~83% | ~70% | 87.7% |
Data: public docs / as of 2026 Q1 / Benchmarks from third-party test sets