AI RESEARCHarXiv9d ago

Invalidation Contracts for Cross-Episode Agent Memory

Michael Wu · Arquimedes Canedo

arXiv:2609.00243v1Artificial Intelligence

WHAT THIS RESEARCH IS ABOUT

LLM agents that cache API error-recovery suggestions save tokens and calls across episodes, but server-side data drift makes these cached fixes silently fail. The conventional workaround of re-deriving fixes on every episode wipes out these efficiency gains.

To address this, researchers introduce invalidation contracts, a protocol layer that equips recovery suggestions with version stamps and cacheability hints. This mechanism allows clients to selectively evict stale entries and retain valid ones without relying on trial and error.

In an evaluation across seven models and approximately 9,400 episodes, row-level invalidation achieved perfect eviction precision and cut baseline token costs by 29% to 33% on four models at the cost of a 15% payload increase. However, the researchers found that realized savings are split between protocol validity, which is completely deterministic, and model compliance, which varied sharply because certain models refuse fixes that introduce new fields.

AI-assisted summary of the paper abstract.

KEY POINTS

01
Invalidation contracts
The proposed protocol attaches version stamps and cacheability hints to error-recovery suggestions so agents can evict stale memory without trial and error.
02
Planner compliance variation
While validity is deterministic across vendors, first-try compliance ranged from 100% on Claude Haiku 4.5 down to 11% or below on Claude Sonnet 5 due to input-schema conservatism.
03
Granularity impact
Row-level invalidation achieved perfect eviction precision and recovered 29% to 33% of baseline token costs on four of seven models, whereas table-level invalidation dropped post-drift first-try rates to 0% on five models.
04
Low overhead
The contract adds a 15% payload overhead and experienced zero contract failures across roughly 9,400 evaluated episodes.

Abstract

LLM agents that cache recovery suggestions from API errors can skip re-derivation in later episodes, spending fewer tokens and fewer model calls on constraints they have already learned. Server-side data drift turns those cached fixes into silent failures, and the usual remedy, re-deriving on every episode, gives the savings back. We introduce invalidation contracts, a protocol layer that attaches version stamps and cacheability hints to every recovery suggestion so the client can evict stale entries without trial and error, and keep the rest. The contract decomposes realized savings into two independent factors: validity, the fraction of cached suggestions that remain correct after a drift event, and compliance, the fraction the planner applies on the first attempt. Validity depends only on the protocol and is vendor-independent. Compliance depends on the planner model: identical wire bytes yield 100% first-try compliance on Claude Haiku 4.5 and 11% or below on Claude Sonnet 5, which exhibits input-schema conservatism, refusing fixes that add fields the original request did not contain. We evaluate across seven models, three serving paths, two domains, and approximately 9,400 episodes. Row-level invalidation raises compliance by 0 to 66.7 percentage points across the seven models, 55.6 to 66.7 on three, and recovers 29-33% of baseline token cost on four of seven models, while table-level invalidation destroys co-located entries and drops post-drift first-try rates to 0% on five of seven. Eviction precision is 1.00 at row granularity on every model under the row-level oracle of Section 4.1. The contract adds 15% to response payload. Version-stamp validity is deterministic by construction and produced identical results across every model and serving path, with zero contract failures in the entire evaluation.

Discussion · 0

Original paper on arXiv