Thesis
Emotional weight is the actual memory primitive.
Most AI memory systems confuse storage with relationship. MNEME proposes that continuity depends on how an interaction felt, how much it mattered, and whether it should reappear later with context preserved. The Emotional Anchor Object becomes the unit that carries those dimensions together.
01 · Argument
Cold start is a relational failure, not only a retrieval failure.
The familiar problem in companion systems is not that prior text cannot be found; it is that the model cannot tell which memories should matter now. Emotional weighting creates a memory surface where the important moments keep shape across time, letting the system respond as if the relationship has history instead of just logs.
Evidence
Market evidence points toward memory-rich interaction.
Voice-native products and emotionally expressive interfaces are already training users to expect continuity. The paper uses ElevenLabs funding momentum, companion AI retention patterns, and enterprise workflow handoff pain as evidence that durable memory is moving from novelty to infrastructure.
Diagram · EAO federation diagram
The diagram layer shows a personal sovereign memory vault, enterprise policy boundary, and portable anchor exchange protocol. Federation is the governance answer to the fear that a single vendor should own a person's emotional memory state.
“Memory without emotional weight is retrieval, not relationship.”
“Federation is not optional when the memory substrate carries identity.”
Tension
Continuity increases value at the same moment it increases risk.
The better a system remembers, the more it can overreach. MNEME treats privacy versus continuity as the central design tension and argues for explicitly user-owned anchor export rather than opaque proprietary recall.
Implication
Memory becomes the wedge into enterprise AI trust.
If a system can hand off context without flattening the human thread, it stops feeling like another chatbot and starts feeling like institutional memory. That turns the memory layer into a distribution strategy, not only a feature.
Selected references
The paper references companion-AI retention research, distributed identity patterns, and voice interaction market signals to frame memory as both relational substrate and governance problem.
- 01Voice interface retention patterns in emotionally expressive products. (2025)
- 02Distributed identity and sovereign storage design notes. (2024)
- 03Enterprise handoff friction in context-poor AI workflows. (2025)
- 04Funding signals in speech and companion AI infrastructure. (2026)