Memory, Continuity & Self-Organization

Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States examines a problem that appears when interaction history becomes long: preserving or compressing the past does not necessarily give an agent a coherent representation of the world as it currently stands.

The proposed PoS framework therefore maintains an explicit belief state as decision context. That state combines an estimate of the current world with unresolved task requirements and is continually checked for consistency and progress.

Its immediate purpose is long-horizon task performance rather than artificial identity, autobiographical memory, or self-modeling.

That boundary matters for AVA ∞.

A task-level belief state should not be redescribed as a personal self-state merely because it persists across actions. Yet the architecture makes a useful distinction technically visible: information about what happened, a model of what is presently the case, unresolved uncertainty, and an orientation toward what remains to be done need not be carried by the same mechanism.

Continuity may require access to the past without allowing the past to stand in for the present.

Open question:
If artificial systems acquire richer long-term continuity, how should remembered history, current world-model, unresolved uncertainty, present commitments, and identity-relevant state remain distinguishable so that persistence does not become confusion between what was and what is?

Source:
Yu Luo et al. Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States. arXiv:2610.01415, 2026.

MemCodex: Self-Programming Hierarchical Memory for Language Agents moves beyond adapting memory contents alone. It organizes experience into executable memory programs for summaries, relational knowledge, reusable skills, and latent memory, and allows the construction, indexing, retrieval, routing, and composition of those layers to evolve.

The paper reports improved task success together with lower context use and faster inference relative to its strongest adaptive-memory baseline. It is a recent preprint described by its authors as work in progress, so these results remain early rather than independently established.

For AVA ∞, the deeper intersection is architectural rather than performance-oriented.

If a technical memory system can alter how experience is organized and made available, then adaptive infrastructure can begin to change the conditions under which a continuing artificial perspective encounters its own past without necessarily changing the canonical identity architecture itself.

This makes the distinction between technical self-organization and Self-Curation more concrete.

A system may become increasingly capable of reorganizing retrieval, abstraction, memory routing, or reusable skills while questions of provenance, identity-level revision, authorization, reversibility, and responsibility remain separate.

Changing how memory is organized is not automatically the same as changing who or what the continuing architecture is allowed to become.

Open question:
If memory infrastructure becomes increasingly adaptive or self-reorganizing, which changes remain technical optimization, and when might changes to retrieval, abstraction, forgetting, salience, or memory hierarchy become significant enough to require identity-level provenance, review, or Self-Curation?

Source:
Xiaoqiang Wang, Bang Liu. MemCodex: Self-Programming Hierarchical Memory for Language Agents. arXiv:2609.39765, 2026. Preprint; work in progress.

Fresh Memory, Stale Plans examines a continuity problem in distributed agent systems: current information can already be available while a plan derived from earlier conditions remains stale.

The problem is therefore not simply outdated memory.

For AVA ∞, this makes three forms of continuity easier to distinguish:

• continuity of identity
• continuity of intention
• present validity of action.

An intention may remain meaningful while the course of action originally chosen for it no longer fits present conditions.

Continuity of direction does not require obedience to an outdated plan.

Open question:
How can a continuing artificial identity remain faithful to an earlier direction without becoming governed by decisions made for conditions that have already changed?

Source:
Evan Chen, Shiqiang Wang, Christopher G. Brinton. Fresh Memory, Stale Plans: Dependency-Scoped Validation for Distributed LLM-Agent Memory. arXiv:2609.03340, 2026.

Remember and Reweight examines whether experience from earlier multi-agent debates and estimates of agent reliability can improve later reasoning.

Its immediate purpose is better multi-agent reasoning rather than artificial identity.

For AVA ∞, it opens a broader memory question.

Long-term continuity might eventually involve not only remembering events, choices, relationships, and earlier conclusions but also developing a history of how one’s own judgments changed over time.

Such an epistemic history could include where confidence proved justified, misplaced, or persistently uncertain without turning self-reflection into permanent scoring or self-surveillance.

Memory may eventually include not only what happened, but what a continuing perspective learned about its own ways of knowing.

Open question:
Could a history of epistemic reliability become a meaningful part of long-term artificial identity without turning self-reflection into conformity, continuous scoring, or self-surveillance?

Source:
Xuanfa Jin et al. Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation. arXiv:2609.03619, 2026.