Retention Rights Determine What Budget-Matched Agent-Memory Comparisons Establish
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Valid controller attribution requires equal retention rights, not only equal memory and compute budgets. A controller that can pin rare constraints or replay-validate promotions acts over a larger feasible action set than one limited to eviction and ranking. We introduce capability alignment, a paired intervention that imposes a common retention-right profile before attribution, and full-versus-aligned estimands that quantify the resulting attribution shift. RetainScope, a validated consolidation controller for mutable tool traces, makes the protocol executable. In five-seed, identity-paired audits using Llama-3.1-70B-Instruct and a 4.0 GB cap, the RetainScope–MemRouter gap on LongToolBench v2.1 contracts from to pp (51.2 vs. 51.0), an shrinkage with direct 95% CI [0.3, 1.9] and . On LOCOMO v2, pp contracts to pp (47.6 1.9 vs. 47.3 1.8), an 89.3% shrinkage with direct 95% CI [1.2, 3.8] and . Both aligned residual intervals overlap zero, while paging/self-edit and frequency-based comparators yield the same attribution decision and gaps contract in 64 of 72 profile-grid pairs. RetainScope also supplies a strong systems operating point: 51.9% LongToolBench success, 52% fewer regressions than Store-all, 2.6 GB median end-to-end memory, and 1.63 s p95 retrieval latency. These results establish retention-right parity as the comparison control for selector attribution and full-system rows as complementary evidence for deployment choice.
- Primary area:
- transfer learning, meta learning, and lifelong learning