Cartographic Memory: An Impossibility Framework and Distortion Ledger for Agent Memory Consolidation
Year: 2026
Type: preprint
Abstract: Memory consolidation in AI agents compresses episodic traces into compact representations, and every such compression is lossy. The field has measured the resulting degradation: continuous LLM-based consolidation can reduce memory utility below the no-memory baseline, and iterative summarization distorts information progressively. Yet no deployed memory system declares which properties its consolidation preserves, and none accounts for where distortion accumulates. We argue that this problem has a structural parallel in cartography, where Gauss's Theorema Egregium (1827) proves that no flat map of a curved surface can preserve all metric properties, and Tissot's indicatrix (1881) provides a per-region instrument for measuring what each projection distorts. We introduce a framework that reads memory consolidation as a projection choice among mutually exclusive invariants and equips each consolidation step with a recomputable per-region distortion ledger. We test three ledger laws (conservation, monotonicity, composition) across three regimes on public-domain books: linear projection (12 books), extractive selection (12 books), and LLM-based abstractive summarization (8 books). In the linear and extractive regimes, all laws hold exactly. In the abstractive regime under a vocabulary-bound tf-idf instrument, global fidelity monotonicity holds on all 8 books; per-region monotonicity fails in 6 of 8; and chaining two passes always loses more fidelity than one deeper pass. A sentence-embedding robustness check (all-MiniLM-L6-v2) reveals that the results are instrument-dependent: global monotonicity fails on 1 of 8 books (Hume's fidelity rises across passes, suggesting parametric regeneration rather than compression), per-region violations increase from 11 to 43 of 128 comparisons, and the composition gap reverses sign on 2 of 8 books. The instrument dependence is itself a finding: different similarity functions expose different failure modes, and the per-region ledger catches what global measures hide.
Author: Green, Phillip Austin
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225984607
