MutMem: Cryptographic Protocol for Authorized Memory Mutation in Persistent Agent Memory
A new cryptographic protocol called MutMem has been introduced to tackle the attribution issue in persistent agent memory systems. You can find more details in a paper on arXiv (2608.02843). This protocol works with the HOM-AIMOS memory engine and allows memory to adjust as new information impacts earlier data, all while ensuring that changes are authorized and verifiable to prevent database tampering. It maintains memory records indefinitely, logging both positive and negative evidence. Significant changes are tracked as transitions approved by a housekeeper, with details linking to various elements, including SHA-256 commitments. The system also verifies data integrity using Ed25519. The study shows that HOM-AIMOS effectively answers 459 out of 500 questions, highlighting its importance for trustworthy AI memory systems.
Key facts
- MutMem is a protocol for authorized mutation in persistent agent memory.
- It is implemented in HOM-AIMOS, a persistent agent-memory engine.
- The protocol addresses the attribution problem in mutable retrieval weights.
- It records signed positive and negative outcome evidence without age-based expiry.
- Each weight change is committed as a housekeeper-authorized transition.
- Transitions bind provenance nodes, signer epochs, quantized weights, no-fork predecessors, and SHA-256 commitments.
- Ed25519 verification is used in both the database writer and a portable verifier.
- Poison-likely content is retained with signed, revisable labels.
- HOM-AIMOS answers 459/500 questions in evaluation.
- The paper is available on arXiv with ID 2608.02843.
Entities
Institutions
- arXiv