Immutable Anchor Bank: A Standalone Memory Protocol for Non-Overwriting, Conflict-Preserving External Memory
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The manuscript presents the Immutable Anchor Bank as a standalone external-memory protocol for systems that need to preserve exact remembered objects rather than only compressed semantic traces. Its central claim is that the novelty is not the mere existence of external memory, but the structural obligations imposed on that memory: stored items are treated as persistent anchors with stable identity, preserved content, provenance, and typed relations. The document argues that exact memory should follow different update rules from compressed latent state, especially when overwrite, averaging, or blending would destroy distinctions that later reasoning may need. At the conceptual level, the protocol is built around a small set of core invariants. Existing anchors must not be overwritten; distinct anchors must not be semantically blended; updates must follow an append-or-link discipline; contradiction must be preserved rather than erased; and relation structure must remain explicit. The document defines four main relation types—duplicate, alias, variant, and conflict—and argues that these relations are not decorative metadata, but part of the semantic contract of the memory regime. Their role is to let the system preserve difference without collapsing everything into flat similarity-based storage. Operationally, the manuscript specifies both write semantics and read semantics. On the write side, candidate memory objects are proposed by an upstream writer, ranked under a fixed write budget, compared against an existing anchor neighborhood, and then either added as new anchors or linked structurally as duplicate, alias, variant, or conflict. On the read side, the bank is explicitly not treated as a plain nearest-neighbor store: retrieval is staged, may use multiple probes, and is intended to surface not just single items but also relation-aware neighborhoods such as variants, aliases, or conflict sets when needed. The document also positions the protocol against other memory regimes. It argues that the Immutable Anchor Bank should not be framed as a replacement for all memory, but as a protected memory regime for cases where object identity, revision history, provenance, and contradiction matter. It is contrasted with mutable caches, overwrite-based key-value stores, averaged latent memory, summary memory, and flat vector stores. The intended claim is narrow and disciplined: this is a candidate regime for preserving recoverable distinction, not a universal solution to long-context memory. A major portion of the manuscript is devoted to scientific discipline rather than hype. It identifies the main reasons the protocol might be useful—precision preservation, variant preservation, conflict preservation, auditability, and better failure diagnosis—but also makes clear that these are hypotheses, not established victories. The document lists serious failure modes such as anchor inflation, duplicate explosion, relation noise, false conflict proliferation, indexing overhead, superficial self-writing, and the possibility that the host system may simply ignore the anchor branch. It also stresses that teacher-assisted bootstrap may help implementation, but should not be confused with proof of the protocol’s independent value. The validation section is one of the document’s strongest parts. It proposes a staged evaluation program that separates structural correctness, practical survivability, and marginal value. In other words, the bank must first prove that it really obeys its own invariants, then show that it remains usable under realistic constraints and scaffold reduction, and only after that be judged on whether it adds value beyond simpler baselines. The document is explicit that the key question is not whether a model with anchors can do something interesting, but whether the immutable-anchor branch produces enough marginal benefit at acceptable cost to justify itself relative to strong non-anchor controls. Finally, the manuscript includes a substantial mathematical appendix. That appendix formalizes anchors, banks, update operators, read operators, and the protocol’s invariants; proves preservation and monotonicity results; derives growth, storage, and complexity bounds; formalizes structured retrieval and multi-probe recovery; analyzes robustness under approximate relation classification; provides counterexamples against overwrite and blending regimes; and, in the later math batches, adds stronger degradation bounds, separation/tradeoff statements, and a map distinguishing what is already proved, what is conditional on assumptions, and what remains open. This gives the document not just a conceptual narrative, but the beginnings of a real theorem-backed foundation.
Fields of Study & Subjects
Advanced Database Systems and Queries
Cognitive Computing and Networks
Advanced Data Storage Technologies
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN
- Document type
- Preprint