Felipe Maya Muniz of AletheionAGI introduces ASM-CM (Aletheion Compact Memory Model), an experimental architecture that separates memory from language to allow agent history to grow without increasing active computational memory.
- The system maintains a compact active state and retrieves only relevant associations, sending minimal context to a local or remote LLM like Qwen3 8B.
- In the "Diamond Promise" experiment, ASM-CM passed 15 out of 15 cases, surviving up to 10,000 distractors per character and process restarts.
- Under a frozen 32K streaming protocol, it achieved 100% MQAR associative-retrieval accuracy with approximately 140 KiB of retained state per stream.
- Endurance protocols for 1-hour and 6-hour persistence have passed, while the 24-hour test is currently being finalized.
The architecture aims to reduce input token costs and latency by avoiding the repeated transmission of full conversation histories to external models.