The Memory Problem Holding Back AI Security Agents

Truly autonomous SOC agents need to learn from incidents, reconcile contradictory information and retrieve relevant experiences – they don’t, yet.

Security vendors are rapidly incorporating agentic AI capabilities into SIEM, XDR and SOC platforms. Microsoft, Splunk, CrowdStrike and Palo Alto Networks are using AI agents to automate triage, investigation, threat detection, incident enrichment and even initial response actions. Yet a fundamental limitation remains: without persistent, auditable memory, many of these systems are still highly sophisticated workflows rather than truly learning security agents.

The market is moving faster than many security operations centres (SOCs) can adapt organizationally. Agentic AI is no longer a future concept. Microsoft introduced Security Copilot agents for phishing, DLP and insider-risk triage in 2025, while Splunk expanded its portfolio in 2026 with role-specific agents for detection engineering, SOP execution, malware analysis and response automation. CrowdStrike and Palo Alto Networks have similarly positioned their platforms as foundations for an "agentic SOC".

The pressure driving this shift is immense. Attackers are increasingly using AI at machine speed, shrinking the window for defenders from weeks to hours "“or even seconds. As a result, manual triage and siloed tools no longer scale effectively.

A Lack Of Persistence

At the heart of the debate is memory.

For AI agents, memory is the ability to retain, organize and selectively retrieve information across interactions. It is this capability that turns a stateless text generator into an adaptive agent. In a security operations context, that means remembering not only an individual alert, but also previous incidents, verified false positives, known exceptions and the effectiveness of past countermeasures.

Evidence that such capabilities remain immature can be seen in vendor messaging. Splunk talks about moving from "isolated alerts" to "entity-aware investigations". The language itself highlights the challenge: if alarms are treated as individual events rather than part of a continuing narrative involving users, assets and identities, the system lacks meaningful context. An agent may perform well during a single investigation but effectively starts again from scratch during the next.

RAG Helps, But It Is Not A Memory System

The popularity of retrieval-augmented generation (RAG) has led some to assume the problem is already solved. Research suggests otherwise.

While retrieval systems form an important component of AI memory architectures, they represent only one part of a broader cycle involving writing, managing and retrieving knowledge. Without mechanisms for consolidation, prioritization, contradiction handling and deliberate forgetting, organizations end up with a more capable search function rather than a true long-term memory system.

In a SOC, this distinction matters. An agent must identify which previous incidents, identities, assets or policy deviations are relevant to the current situation. This is why vendors are investing heavily in context-rich data models.

CrowdStrike's Enterprise Graph, for example, aims to create a connected view of telemetry across endpoints, identities, cloud environments and third-party tools. Splunk's Entity Analytics takes a similar approach. These developments strengthen context, but they do not yet amount to genuine memory in the cognitive sense.

What A Real SOC Memory Would Need

A genuinely autonomous security agent would require three forms of memory. First, episodic memory to retain information about completed incidents, investigation paths and successful remediation actions. Second, semantic memory for relatively static knowledge such as asset ownership, identities, policies and business processes. Third, procedural memory for playbooks, escalation paths and standard operating procedures. Modern agent-memory research increasingly reflects this three-part structure.

Most current platforms primarily address the procedural layer and, to some extent, the semantic layer. Microsoft’s Phishing Triage Agent learns through administrator feedback, Splunk’s SOP Agent translates operating procedures into response plans, while Palo Alto Networks and CrowdStrike have introduced specialized agents covering threat intelligence, email investigations, malware analysis and workflow orchestration. However, long-term episodic memory remains largely absent.

Progress Is Real But Uneven

The industry is moving towards more sophisticated approaches. Three trends are particularly visible.

The first is the development of richer contextual data layers that link entities, telemetry and historical events. The second is the growing use of feedback loops, with vendors claiming their systems improve through analyst input. The third is the increasing incorporation of procedural knowledge directly into agentic workflows.

These developments are important but should not be confused with autonomous learning. Long-term knowledge consolidation across multiple incidents and sessions remains largely unsolved. Researchers continue to identify open challenges around contradiction handling, latency constraints, privacy governance, trustworthy reflection and what they describe as "learned forgetting".

A Governance Issue As Much As A Technology Problem

As agents move closer to operational decision-making, memory becomes a governance concern as well as a technical one.

Palo Alto emphasizes role-based access controls, human approval for high-impact actions and full auditability. CrowdStrike promotes "bounded autonomy" with expert-defined guardrails. Splunk explicitly describes its vision as "agentic-led and human-governed", while Microsoft focuses on explainable decisions and improvement through administrator feedback rather than uncontrolled self-optimization.

In Europe, the issue also has GDPR implications. If agents are expected to learn persistently from identity data, DLP incidents, insider-risk signals and email content, questions about data ownership, privacy and tenant boundaries become unavoidable. Researchers regard privacy governance as an unresolved engineering challenge, while vendors simultaneously highlight the value derived from vast stores of historical telemetry and operational knowledge.

Automation, Assistance Or Autonomy?

Today's security platforms clearly demonstrate the value of AI agents. They can accelerate triage, enrich investigations, prioritize alerts and automate predefined workflows. Microsoft focuses on phishing and DLP assessments, Splunk on detection engineering and malware analysis, CrowdStrike on agentic detection and response, and Palo Alto on configurable security agents.

Yet these use cases remain bounded. They help sort, explain, enrich and execute. What they do not convincingly demonstrate is sustained organizational memory across months or years of incidents - the sort of experience that seasoned analysts develop through repeated exposure to false positives, policy exceptions and historical incident-response decisions.

This marks the current dividing line between assistance and genuine autonomy. True autonomous SOC agents would need to learn continuously from incidents, reconcile contradictory information, retrieve relevant experiences at the appropriate moment and explain their actions in a fully auditable way. Research suggests these challenges remain unresolved.

The Alternative: Human-Led Security Operations

Not every vendor is pursuing full agent autonomy. Sophos, for example, positions its MDR business around a human-led, AI-assisted model, where AI is used primarily to accelerate analysis while decision-making remains firmly in the hands of analysts. The knowledge base is therefore built around collective operational experience rather than machine memory.

Other MDR-focused providers, including Arctic Wolf, Secureworks and Expel, follow similar approaches. Their emphasis remains on expert analysts supported by AI rather than autonomous systems making independent decisions. In these models, organizational memory resides in people, processes and accumulated operational expertise.

The current generation of AI-powered SOC agents represents a significant advance. They can speed up investigations, reduce analyst workload and improve operational efficiency. However, they still lack the persistent organizational memory needed to transform isolated events into accumulated experience.

For enterprises, the key question is not how much automation a platform promises, but how effectively it manages context, memory, feedback, governance and human oversight. Until long-term memory, auditability and controlled learning mature, human-governed security operations are likely to remain essential.

The SOC of the future will be neither fully autonomous nor entirely human-led, but one where AI provides speed while people retain responsibility, judgement and experience.

This article first appeared on Computing’s sister site Computing Deutschland.