Agent memory is still an unsolved distributed-systems problem: 23 documented failure modes across five areas, from writes that mint false authority to rollbacks that keep approvals they cannot justify

Asked (summary):

What are the unresolved system-level problems in AI agent memory and context management for production agents — persistent state, memory concurrency, context-window costs, context freshness, and memory provenance? Name the sources.

This synthesis draws on 95 sourced findings from a broad live-web scan of research papers, surveys and engineering discussions current through October 2026 — mostly primary arXiv papers, with mirrors and paper-list aggregators de-emphasized. It is a structured synthesis of unresolved issues, not an exhaustive census of the literature. Empirical results are labelled as such; surveys and proposed architectures are labelled separately, and no single benchmark number should be read as universal prevalence.

Where production agent memory breaks: the information pipeline

Information flows top to bottom from observation to action. Each tag is a documented unresolved failure mode at that stage — click or tap a tag for the finding and its source. Colour = problem area.

The five areas interact — fixes in one worsen another

Aggressive compression saves tokens but severs lineage and freshness: a summary that drops a revocation or a rare qualifier is cheaper and wrong. Concurrency multiplies lineage: every derived copy an agent holds must be reachable by a later deletion or revocation. Persistence turns one bad write into a cross-session failure that retrieval keeps resurfacing. And every remedy — verification, provenance capture, event sourcing — adds context, latency and storage cost back onto the window the agent was trying to shrink. The measured safety–utility frontier in authorization laundering (unauthorized use cut from 25.3% to 7.3%, authorized use falling to 53.8%) is the clearest evidence that these are coupled trade-offs, not independent bugs.

What the map cannot say alone

The failure is systemic, not model-sized: swapping to a ~70×-costlier internal LLM is the only intervention that narrows MEME's dependency-propagation gap, and it degrades exact recall while doing so — arxiv.org

Governance is empirically absent, not just immature: no published memory architecture covers all nine governance primitives, and write-gate validation plus post-deletion verification are blind spots everywhere — memorypapers.org

Retrieval, not generation, dominates long-horizon memory failure in one benchmark — over 93% of failures originated in the memory system's retrieval rather than the answering model (aggregator-sourced; read as indicative) — github.com

ACON's 26–54% peak-token reduction is progress, not closure: the compressor is itself an LLM call, it invalidates KV-cache reuse, and history compression rarely reduced net cost in the authors' own measurements — arxiv.org

Five areas, thirteen unresolved problems

Area	Unresolved system problem	Why current approaches fall short	Production consequence	Partial direction	Source

Structured synthesis of 95 sourced findings (30 on persistent state, 25 on concurrency and shared state, 40 on context cost, freshness and provenance) gathered by broad live-web search across papers, surveys, framework docs and engineering discussions, current through 2026-10-08. Duplicate versions of the same paper were merged and aggregator rows de-emphasized in favour of primary papers. Percentages are each paper's own reported benchmark results, not population prevalence. The pipeline map shows 23 representative failure modes; the table condenses to 13 problem rows for space. This is not an exhaustive census of the literature.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.