AIOps Time-Series Memory Research Data

Date: 2026-06-13 Branch: codex/aiops-time-series-memory-research

This file records the comparison data behind the AIOps research direction in docs/research/time_series_agent_memory_edge.md. It is intentionally structured as a research data log rather than a narrative memo.

Action-outcome engine procedure: docs/research/action_outcome_memory_engine_plan.md

Research Question

Can a time-series-native incident memory layer improve AIOps agents compared with text-only incident RAG or dashboard-centric observability?

Primary hypothesis:

Event-time-aligned incident memory that combines metrics, logs, traces, topology, deploy events, postmortems, and action outcomes will improve similar incident retrieval, root-cause ranking, and runbook recommendation for AIOps agents.

Commercial hypothesis:

The first useful product is read-only Similar Incident Retrieval plus Evidence Pack APIs for AI SRE workflows. Guarded remediation can come later after trust, policy, and audit controls are proven.

Market Data

Market estimates vary widely because reports define “AIOps” differently. This is useful signal: the category is active, but product positioning must be precise.

Source	Segment	Base Value	Forecast Value	CAGR	Notes	URL
Mordor Intelligence	AIOps	USD 18.95B in 2026	USD 37.79B in 2031	14.8%	Mentions ML correlation engines and lower MTTR as demand drivers.	https://www.mordorintelligence.com/industry-reports/aiops-market
Fortune Business Insights	AIOps	USD 2.23B in 2025; USD 2.67B in 2026	USD 11.8B in 2034	20.40%	Much smaller definition than Mordor; still shows strong growth.	https://www.fortunebusinessinsights.com/aiops-market-109984
Future Market Insights	AIOps platform	USD 15.8B in 2025; USD 19.8B by 2026-end	USD 187.2B in 2036	25.2%	Aggressive long-range estimate; use as upside case only.	https://www.futuremarketinsights.com/reports/aiops-platform-market
Market Research Future	AIOps platform	USD 10.52B in 2024; USD 12.43B in 2025	USD 66.2B in 2035	18.2%	Mid-range AIOps platform estimate.	https://www.marketresearchfuture.com/reports/aiops-platform-market-11745
Coherent Market Insights	AIOps platform	USD 14.69B in 2026	USD 68.88B in 2033	24.7%	Supports a high-growth platform view.	https://www.coherentmarketinsights.com/market-insight/aiops-platform-market-2073
MarketsandMarkets	AIOps platform	USD 11.7B in 2023	USD 32.4B in 2028	22.7%	Older but widely cited platform estimate.	https://www.marketsandmarkets.com/Market-Reports/aiops-platform-market-105974848.html
Mordor Intelligence	Autonomous IT operations	USD 17.28B in 2026	USD 38.34B in 2031	17.28%	Adjacent category directly tied to self-healing workflows.	https://www.mordorintelligence.com/industry-reports/autonomous-it-operations-market
Mordor Intelligence	Observability	USD 3.35B in 2026	USD 6.93B in 2031	15.62%	Mentions AI-driven and edge-centric workloads as demand catalysts.	https://www.mordorintelligence.com/industry-reports/observability-market
Credence Research	Incident management software	USD 7.215B in 2024	USD 15.579B in 2032	10.1%	Adjacent buyer budget for incident workflows.	https://www.credenceresearch.com/report/incidence-management-software-market
MarketsandMarkets	Predictive maintenance	USD 13.89B in 2026	USD 23.79B in 2031	11.4%	Relevant for industrial AIOps and physical AI edge expansion.	https://www.marketsandmarkets.com/Market-Reports/operational-predictive-maintenance-market-8656856.html

Product Comparison

The market is moving toward AI SRE agents, agentic observability, and automated incident workflows. The gap for ZeptoDB is a self-hosted or edge-deployable time-series memory plane that can feed these agents with structured evidence.

Vendor	Product / Capability	Stated Capabilities	Data Modalities	Action / Automation	Evidence / Audit Signal	ZeptoDB Opportunity	URL
Datadog	Bits Investigation	AI SRE agent grounded in thousands of incidents; claims root-cause identification 90% faster.	Datadog telemetry and incident data.	Troubleshooting and root-cause investigation.	Product page emphasizes root-cause speed.	Customers may want customer-owned incident memory and exportable evidence packs.	https://www.datadoghq.com/product/ai/bits-investigation/
Datadog	Bits AI SRE	Autonomous alert investigations, synthetic API test investigations, APM graph-triggered investigations, recommended actions, Slack/Jira triage.	Metrics, logs, traces, dashboards, changes; later blog adds source code, RUM, DB monitoring, network path, profiler, events.	Recommended actions, ticket/chat actions, code-fix handoff through Bits AI Dev Agent.	Agent Trace view shows tools called, data queried, and intermediate analysis.	ZeptoDB can compete below the UI: long-term incident memory, third-party telemetry, edge/site deployments.	https://www.datadoghq.com/blog/bits-ai-sre/
Datadog	Bits AI SRE deeper reasoning update	Claims about 2x faster investigations, roughly 3-4 minutes depending on complexity; expanded data sources; third-party integrations in preview.	Full-stack Datadog plus selected third-party sources.	Triage, assignment, automation integrations.	Agent Trace and hypothesis tree.	A ZeptoDB memory layer should expose comparable agent traces and evidence provenance.	https://www.datadoghq.com/blog/bits-ai-sre-deeper-reasoning/
Dynatrace	Dynatrace Intelligence	Fuses deterministic insights with agentic action for autonomous prevention, remediation, and optimization.	Dynatrace observability and platform data.	Autonomous prevention, remediation, optimization.	Deterministic AI plus policies and guardrails.	ZeptoDB should not claim pure LLM RCA; it should combine deterministic temporal evidence with agent reasoning.	https://www.dynatrace.com/platform/artificial-intelligence/
Dynatrace	Agentic Dynatrace Assist	Uses skills, reference files, and MCP tools to query Dynatrace environment; user permissions scope results.	Dynatrace platform data and official support resources.	Can call MCP tools subject to permissions.	User-based access, permissions, PII prompt blocking behavior.	ZeptoDB needs scoped tool access, tenant isolation, and audit logs for agent queries.	https://docs.dynatrace.com/docs/dynatrace-intelligence/agentic-and-generative-ai/agentic-and-generative-ai-getting-started
Splunk	Observability Troubleshooting Agent	Automatically correlates metrics, events, logs, and traces; surfaces likely root causes and next steps.	Metrics, events, logs, traces.	Root-cause support and next-step recommendations.	Built into Observability Cloud alert workflows.	ZeptoDB can target teams that need portable/self-hosted incident memory across Splunk and non-Splunk data.	https://www.splunk.com/en_us/blog/observability/splunk-observability-ai-agent-monitoring-innovations.html
Splunk	AI troubleshooting agent	Automatically triggers RCA for selected APM service and Kubernetes infrastructure alerts.	APM services and Kubernetes infrastructure monitoring.	Suspected root-cause display and remediation plan.	Alert-linked RCA flow.	ZeptoDB can start with Kubernetes + OpenTelemetry incident memory as a focused MVP.	https://help.splunk.com/en/splunk-observability-cloud/create-alerts-detectors-and-service-level-objectives/create-alerts-and-detectors/ai-troubleshooting-agent-and-remediation-plan
PagerDuty	AIOps	Reduces alert noise, improves incident visibility, triage, and removes repetitive work.	Alerts, incident events, change events, historical incidents.	Event grouping, triage, workflow automation.	Incident lifecycle context.	PagerDuty is a strong integration source for incident labels, timelines, and action outcomes.	https://www.pagerduty.com/platform/aiops/
PagerDuty	Event Intelligence	Claims filtering up to 98% of noise; surfaces relevant incidents, recent changes, and likely origin points.	Alerts, related incidents, recent changes.	Alert grouping and origin suggestions.	”What happened, when” situational awareness.	ZeptoDB should store the richer raw windows behind PagerDuty events.	https://www.pagerduty.com/platform/aiops/event-intelligence/
PagerDuty	SRE Agent	Spring 2026 release says PagerDuty SRE Agent investigates and resolves complex incidents at enterprise scale.	PagerDuty Operations Cloud context.	Investigation and resolution.	Enterprise incident workflow context.	ZeptoDB can provide the time-series evidence layer that incident platforms can call.	https://www.pagerduty.com/newsroom/pagerduty-operations-cloud-spring-2026-release/
Grafana	Assistant Investigations	Uses specialized agents over metrics, logs, traces, and profiles; internal case found root cause in 8 minutes vs 28 minutes manually.	Metrics, logs, traces, profiles, query plans, MCP integrations.	Recommendations for mitigation/remediation.	Evidence trail, hypotheses, confidence scores.	This validates a multi-agent evidence-pack design; ZeptoDB can supply the memory backend.	https://grafana.com/blog/a-tale-of-two-incident-responses-how-our-ai-assist-helped-us-find-the-cause-3-5x-faster/
Grafana	AI-powered Observability	Root-cause analysis, dashboard creation, query suggestions, SRE agent, cost optimization, AI Observability preview.	Grafana Cloud observability data and AI app telemetry.	Troubleshooting and workflow assistance.	Grafana Assistant embedded workflows.	ZeptoDB should integrate through OpenTelemetry/Grafana rather than require UI replacement.	https://grafana.com/products/cloud/ai-observability/
ServiceNow	ITOM AIOps agentic workflows	Network of AIOps agents gathers real-time data from other tools and presents alert/incident context conversationally.	Alerts, incidents, external tool data.	Agentic workflows for incident/outage response.	Platform governance context.	ServiceNow is a workflow and ticket system integration for action outcomes and postmortems.	https://www.servicenow.com/community/itom-blog/revolutionizing-it-operations-with-ai-agents/ba-p/3271551
OpenTelemetry	AI agent observability conventions	Standardized traces, metrics, and logs for AI agent frameworks; GenAI semantic conventions.	Agent traces, model calls, token usage, latency, logs, metrics.	Telemetry standard, not an agent product.	Vendor-neutral instrumentation.	ZeptoDB should ingest OTLP and preserve agent traces as first-class incident evidence.	https://opentelemetry.io/blog/2025/ai-agent-observability/

Benchmark And Paper Data

Source	Type	Data / Environment	Task	Reported Scale	Why It Matters For ZeptoDB	URL
AIOpsLab	Benchmark framework	Deploys microservice cloud environments, injects faults, generates workloads, exports telemetry, and exposes agent interfaces.	End-to-end AIOps agent evaluation across incident lifecycle.	Framework, not just static dataset.	Best fit for evaluating agentic workflows with ZeptoDB as memory backend.	https://arxiv.org/abs/2501.06706
RCAEval	Open-source benchmark	Multiple microservice RCA datasets and baselines.	Root-cause analysis.	9 datasets, 735 real failure cases, 15 reproducible baselines.	Useful baseline suite for comparing time-series memory retrieval against existing RCA methods.	https://github.com/phamquiluan/RCAEval
OpenRCA	Benchmark dataset	Real software operating scenarios with telemetry.	LLM root-cause analysis.	335 failure cases, 3 heterogeneous software systems, over 68 GB telemetry.	Direct benchmark for testing whether LLMs need structured time-series evidence.	https://netman.aiops.org/wp-content/uploads/2025/05/13411_OpenRCA_Can_Large_Langua.pdf
CCF AIOPS 2025 RCA Challenge dataset	Challenge dataset	Extended from Google HipsterShop microservice system; includes trace, metric, and log data.	Root-cause analysis.	Public challenge dataset referenced by 2026 RCA paper.	Good for controlled experiments over metrics/logs/traces.	https://arxiv.org/html/2602.08804v1
Survey of AIOps in the Era of Large Language Models	Survey	Literature map for LLM-based incident management, log analysis, RCA, mitigation, postmortems, and QA.	Survey / taxonomy.	Broad paper survey.	Helps position this research against text-only incident reports and LLM-only RCA.	https://arxiv.org/html/2507.12472v1
Auditable Graph-Guided RCA for Kubernetes	Method paper	Kubernetes incidents, graph traversal agent, graph-guided RCA.	Auditable RCA.	2026 paper.	Supports combining LLMs with graph/time-series tools instead of free-form diagnosis.	https://arxiv.org/html/2606.08590v1
AIOps “AI Oops”	Safety/security paper	LLM-driven IT operations and RCA attack surface.	Subversion and trust risks.	2025 paper.	Reinforces need for scoped tools, evidence packs, audit trails, and safe action gates.	https://arxiv.org/html/2508.06394v2
TrioXpert	Automated incident management framework	Incident detection, root-cause localization, mitigation, and post-incident stages.	Incident lifecycle automation.	2025 paper.	Useful lifecycle model for mapping memory objects to actions and outcomes.	https://arxiv.org/html/2506.10043v1
Cloud-OpsBench	Benchmark	Reproducible benchmark for agentic root-cause analysis; references AIOpsLab.	Agentic RCA.	2026 paper.	Candidate evaluation target for interactive agent experiments.	https://arxiv.org/html/2603.00468v1
Uncovering Reasoning Failures in LLMs for Cloud RCA	Analysis paper	Extracts modality-specific alerts from logs, metrics, and traces into unified alerts.	Reasoning failure analysis.	2026 paper.	Supports measuring false-cause rate, not only answer accuracy.	https://arxiv.org/html/2601.22208v1

AIOps Data Modalities To Capture

The research data model must include more than incident text.

Modality	Required Fields	Why It Matters	ZeptoDB Storage Fit
Metrics	metric name, labels, entity, timestamp, value, unit, aggregation window	Detect anomalies, trends, saturation, and recovery.	Strong fit as columnar time-series.
Logs	timestamp, service, host/pod, severity, template id, message hash, raw reference	Capture error modes and rare events.	Store templates/features in ZeptoDB; keep large raw logs by reference if needed.
Traces	trace id, span id, parent id, service, operation, latency, error, timestamp	Localize latency and dependency failures.	Store span summaries and time-window references.
Topology	service graph, dependency edges, pod/node/container, region/cluster	Distinguish text-similar but topology-different incidents.	Store as temporal graph snapshots or edge tables.
Change events	deploy id, commit, config change, feature flag, migration, timestamp	Many incidents are change-triggered.	Strong fit as event stream joined to telemetry windows.
Alerts	alert id, condition, severity, threshold, first seen, resolved time	Incident entry point and evaluation query.	Strong fit as event stream.
Runbook actions	action id, actor, command/tool, target, timestamp, approval, rollback	Required for action-outcome memory.	Strong fit as append-only action log.
Outcomes	success/failure, time to mitigation, time to resolution, side effects	Required for recommendation ranking.	Strong fit as incident metadata.
Postmortems	root cause, contributing factors, lessons, owner, links	Human-labeled ground truth.	Store text embedding refs and link to telemetry.
Agent traces	prompt, tool calls, retrieved evidence, decision, action, token/cost, latency	Audit and repeated-mistake reduction.	Store summarized traces and raw references.

Comparative Experiment Design

The comparison should test whether time-series incident memory adds value over current lower-cost baselines.

Experiment	Baseline A	Baseline B	ZeptoDB Variant	Primary Metrics	Required Data
Similar incident retrieval	Keyword search over postmortems	Vector search over postmortems	Time-series motif + topology + text hybrid retrieval	top-k recall, MRR, retrieval latency	Incident labels, postmortems, telemetry windows, topology
RCA ranking	LLM over alert text only	LLM over alert + logs	LLM with ZeptoDB evidence pack and temporal joins	top-1/top-3 RCA accuracy, false-cause rate, evidence precision	Alerts, metrics, logs, traces, changes, ground truth
Runbook recommendation	Static runbook lookup	LLM runbook Q&A	Action-outcome memory ranked by similar incidents	safe-action rate, successful recommendation rate, rollback rate	Actions, outcomes, runbooks, incident labels
Postmortem generation	LLM over chat transcript	LLM over ticket and logs	LLM with timeline, evidence pack, action log	edit distance, hallucinated claims, evidence coverage	Incident channel, ticket, telemetry, action log
Edge retention	Full cloud retention	Random/sampled retention	Edge anomaly segments + compressed signatures	RCA accuracy per GB, retrieval hit rate, edge storage lifetime	High-frequency telemetry, storage budget, labels
Agent audit	No agent trace	Text-only agent transcript	Structured tool-call/event-time audit log	reproducibility, reviewer acceptance, policy violations	Agent calls, prompts, tool outputs, approvals

MVP Data Contract

The MVP should produce and consume one core record: incident_memory.

Field	Type	Required	Notes
`incident_id`	string	yes	Stable id from PagerDuty/Opsgenie/Jira/internal system.
`tenant_id`	string	yes	Required for SaaS or multi-team deployments.
`environment`	string	yes	prod/staging/site/plant/robot/fleet.
`service`	string	yes	Main affected service or component.
`entity_refs`	array	yes	Host, pod, node, device, robot, region, cluster.
`start_ts`	timestamp	yes	Earliest suspected incident timestamp.
`detect_ts`	timestamp	yes	Alert or human detection time.
`mitigate_ts`	timestamp	no	Service restored but root cause may remain.
`resolve_ts`	timestamp	no	Incident fully resolved.
`event_time_range`	interval	yes	Raw telemetry window by event time.
`ingest_time_range`	interval	yes	Raw telemetry window by ingestion time.
`clock_domain`	enum/string	yes	Needed for edge, robot, replay, and distributed systems.
`symptoms`	array/object	yes	Alerts, metric deviations, log templates, trace anomalies.
`topology_context_ref`	reference	no	Service graph or entity graph snapshot.
`change_context_ref`	reference	no	Deploy/config/feature-flag/migration context.
`anomaly_segment_refs`	array	yes	Pointers into ZeptoDB windows.
`embedding_refs`	array	no	Segment, log, postmortem, runbook embeddings.
`candidate_causes`	array	no	Ranked hypotheses.
`confirmed_root_cause`	string/object	no	Ground truth after postmortem.
`actions_taken`	array	no	Human and agent actions.
`action_outcomes`	array	no	Success/failure, duration, side effects.
`rollback_steps`	array	no	Required for guarded remediation.
`human_notes`	reference	no	Postmortem, ticket, chat transcript.
`agent_trace_ref`	reference	no	Tool calls, evidence, model outputs.
`confidence`	float	no	Model or system confidence.
`evidence_score`	float	no	How strongly retrieved evidence supports the claim.
`safety_score`	float	no	Action safety / reversibility estimate.

ZeptoDB Differentiation Matrix

Capability	Existing AI SRE Vendors	Generic Vector DB	ZeptoDB Time-Series Memory
Human-facing dashboards	Strong	Weak	Not the primary wedge.
Agent-readable evidence packs	Emerging	Weak	Strong opportunity.
Full-fidelity event-time windows	Platform-dependent	Weak	Strong fit.
Similar incident retrieval	Emerging	Text-biased	Hybrid time-series/text/topology retrieval.
Edge/site deployment	Limited for SaaS products	Possible but not telemetry-native	Strong opportunity for gateway-class edge.
Action-outcome memory	Emerging	Weak	Strong if modeled explicitly.
Temporal joins across metrics/logs/traces/changes	Platform-dependent	Weak	Core ZeptoDB opportunity.
Data residency / self-hosting	Varies	Strong	Strong if packaged properly.
Regulated audit trail	Emerging	Weak	Strong if agent_trace and evidence provenance are first-class.
Vendor lock-in avoidance	Weak for SaaS platforms	Strong	Strong if OTLP and incident-tool integrations exist.

Ranked Use Cases

Rank	Use Case	Buyer	Urgency	Data Availability	Automation Risk	Revenue Potential	Recommendation
1	Similar incident retrieval	SRE / platform engineering	High	High	Low	High	Build first.
2	Evidence pack for RCA agents	SRE / incident command	High	Medium-high	Low-medium	High	Build with retrieval MVP.
3	Postmortem-to-memory pipeline	SRE / reliability leadership	Medium	Medium	Low	Medium	Build early to create labels.
4	Runbook recommendation	SRE / operations	High	Medium	Medium	High	Build after action-outcome labels.
5	Edge AIOps recorder	Industrial / robotics / edge ops	Medium-high	Medium	Low	High in verticals	Build after core MVP proves value.
6	Guarded remediation	SRE / IT ops	High	Medium	High	Very high	Later; requires policy and audit.
7	Fully autonomous remediation	CIO / platform leadership	High	Low-medium	Very high	Very high but slow adoption	Research only until trust improves.

High-Risk / Gamechanger Bets

This section compares ideas that are riskier than the read-only MVP but could become category-defining if they work.

Rank	Bet	Field	Gamechanger Thesis	Why Time-Series Memory Matters	Main Risks	ZeptoDB Advantage	Suggested First Proof
1	Closed-loop incident autopilot	AIOps	Move from “AI explains incidents” to “AI safely fixes recurring incidents with evidence, rollback, and audit.”	The agent needs action-outcome memory: what happened, what changed, what it did, whether it worked, and how the system recovered over time.	Production damage, trust, permissions, liability, bad RCA, unsafe automation.	Low-latency incident windows, append-only action logs, temporal joins, evidence packs, and audit traces.	Guarded remediation for 3-5 reversible incident classes such as restart, scale-out, rollback, cache purge, or traffic drain.
2	Edge embodied incident memory / robot black box	Physical AI	Become the memory and replay layer for robots, drones, smart factories, and physical AI fleets.	Physical incidents are temporal and multi-modal: sensor streams, robot state, control actions, operator overrides, environment, and recovery.	Hardware fragmentation, safety certification, huge payloads, edge packaging, robotics adoption cycles.	Existing edge/ROS 2/physical AI docs, aarch64 path, time-series storage, replay orientation.	Robot-local recorder that stores 30-300 seconds of high-fidelity pre/post anomaly telemetry and produces a compact incident signature.
3	Agentic SOC temporal memory	Security / AIOps adjacent	Security operations move at machine speed; agents need unified, high-fidelity timelines to investigate and contain attacks.	Attack chains are temporal graphs across identity, network, endpoint, cloud, and agent actions.	Very high safety risk, adversarial manipulation, compliance, high bar for integrations.	Append-only time-series evidence, action audit, scoped retrieval, self-hosted deployment.	Read-only attack timeline reconstruction plus analyst-approved containment recommendation.
4	Causal remediation simulator	AIOps / Physical AI	Before executing an action, simulate likely blast radius using prior incident memory and system topology.	Counterfactuals need historical temporal neighborhoods, topology, action outcomes, and recovery curves.	Hard causal modeling, false confidence, limited ground truth.	ZeptoDB can store aligned action/outcome trajectories and topology snapshots for retrieval.	For recurring incidents, predict whether restart/rollback/scale-out would reduce error rate within N minutes.
5	Open incident memory protocol	Cross-domain	Define a portable memory object and API that AI agents can use across Datadog, Splunk, Grafana, PagerDuty, ServiceNow, and edge systems.	The protocol must preserve event time, ingest time, evidence links, agent traces, and outcome labels.	Standards are hard, ecosystem adoption is slow, vendors may resist portability.	ZeptoDB can be a reference implementation with OTLP and incident-tool integrations.	Publish `incident_memory` schema plus import/export adapters for OpenTelemetry and PagerDuty.

Gamechanger Ranking Rationale

The best single high-risk bet is Closed-loop incident autopilot. It is risky, but the buyer pain is immediate and the ROI is measurable through MTTR, toil reduction, and repeated-incident elimination. The research should not start with free-form autonomous action. It should start with a narrow loop:

Detect a recurring incident pattern.
Retrieve prior similar incidents and action outcomes.
Produce an evidence pack and ranked remediation.
Execute only reversible, policy-approved actions.
Measure recovery and write the outcome back into memory.

The second best bet is Edge embodied incident memory / robot black box. This has a larger long-term platform upside if physical AI adoption accelerates, but it has a harder route to market because hardware, safety, and robotics deployment cycles are slower than cloud/SRE adoption cycles.

The unifying insight is that both gamechanger bets need the same primitive:

time-series action-outcome memory

This primitive records the temporal situation, the action taken, the policy and approval context, the observed recovery curve, and the long-term outcome. If ZeptoDB owns this layer, it can support AIOps first and physical AI later.

Additional High-Risk Signals

Source	Signal	Relevance	URL
TechRadar OpenClaw coverage	Agentic systems can become unmanaged privileged actors; security requires inventory, permissions, and audit.	Supports treating agents as governed identities before autonomous remediation.	https://www.techradar.com/pro/what-the-openclaw-vulnerability-reveals-about-the-future-of-agentic-ai-security
ITPro / Forrester agentic AI operationalization	Many enterprises are still stuck in pilot because of orchestration, governance, security, and data architecture gaps.	Explains why read-only evidence packs should precede full autonomy.	https://www.itpro.com/technology/artificial-intelligence/most-enterprises-are-still-unprepared-to-operationalize-it-it-leaders-are-bullish-on-agents-but-keeping-falling-at-the-final-hurdle-heres-why
TechRadar agentic SOC	AI-enabled threats increase the need for machine-speed investigation and response.	Makes agentic SOC a high-upside adjacent market for temporal memory.	https://www.techradar.com/pro/security-at-machine-speed-why-the-soc-must-be-rebuilt-for-the-ai-era
Elastic agentic AI SOC	Agentic SOCs promise autonomous prioritization, closed-loop containment, and traceable reasoning.	Reinforces action audit and policy-aligned automation as product requirements.	https://www.elastic.co/security-labs/why-2026-is-the-year-to-upgrade-to-an-agentic-ai-soc
AgentSOC	SOC autonomy must be explainable, risk-aware, and policy-aligned.	Supports graph/time-series evidence over free-form LLM decisions.	https://arxiv.org/html/2604.20134v1
NVIDIA Jetson Thor	Edge robotics compute now reaches 2070 FP4 TFLOPS and 128 GB memory at 40-130 W.	Makes robot-local memory and inference more plausible on edge-class hardware.	https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
Physical Intelligence MEM	Multi-scale embodied memory enables longer robot tasks and in-context correction.	Supports physical AI memory as a real research frontier, not only infrastructure speculation.	https://www.pi.website/research/memory

Adoption Constraints

Constraint	Evidence From Research	Product Implication
Trust and governance slow autonomous adoption	Dynatrace docs emphasize scoped permissions; market commentary says many agent projects remain pilot-stage due to security and governance.	Start read-only. Add approval gates before actions.
AI agents need telemetry, not just chat	Datadog, Splunk, Grafana, and OpenTelemetry all center metrics/logs/traces/tool calls.	Build OTLP and incident integrations before UI polish.
RCA requires multi-modal evidence	Benchmarks and product pages repeatedly mention metrics, logs, traces, topology, changes, and events.	`incident_memory` must be structured and temporal.
Vendor platforms already own dashboards	Datadog, Dynatrace, Splunk, Grafana have mature UIs.	Position ZeptoDB as memory/evidence layer, not dashboard replacement.
Edge deployments need trimmed packages	ZeptoDB docs show edge profiles and resource envelopes, but full image includes heavyweight features.	Create edge AIOps profile later.
Ground truth is hard	RCA datasets rely on failure cases and postmortems; real labels are limited.	Build postmortem-to-memory ingestion early.

Data Gaps To Resolve

Gap	Why It Matters	Proposed Next Step
Real customer incident telemetry	Public benchmarks may not reflect real real operational complexity.	Start with AIOpsLab/RCAEval/OpenRCA, then collect internal replay data.
Precise vendor cost model and attach rates	Needed for business model comparison.	Collect public site pages and public estimates separately.
ZeptoDB OTLP ingest path	Required for easy AIOps adoption.	Audit current ingestion APIs and design OTLP bridge or adapter.
Time-series motif embedding implementation	Needed for retrieval beyond text search.	Prototype segment embedding pipeline over benchmark datasets.
Agent trace schema	Needed for trust and audit.	Align with OpenTelemetry GenAI semantic conventions.
Edge package size/perf	Needed for physical AI and industrial deployments.	Build an edge profile and benchmark on ARM/Jetson-class hardware.

Research Decision

The data supports narrowing the research to:

Event-Time Incident Memory for AI SRE Agents

Source Log

ID	Source	Category	Data Captured
S1	Mordor AIOps market	Market	2026/2031 size, CAGR, MTTR driver.
S2	Fortune Business Insights AIOps	Market	2025/2026/2034 size, CAGR.
S3	Future Market Insights AIOps platform	Market	2025/2026/2036 size, CAGR.
S4	MRFR AIOps platform	Market	2024/2025/2035 size, CAGR.
S5	Coherent AIOps platform	Market	2026/2033 size, CAGR.
S6	MarketsandMarkets AIOps platform	Market	2023/2028 size, CAGR.
S7	Mordor Autonomous IT operations	Market	2026/2031 size, CAGR.
S8	Mordor Observability	Market	2026/2031 size, CAGR.
S9	Datadog Bits Investigation	Product	AI SRE positioning, 90% faster claim.
S10	Datadog Bits AI SRE	Product	Autonomous investigation, actions, code-fix handoff.
S11	Datadog Bits deeper reasoning	Product	2x faster claim, agent trace, expanded data sources.
S12	Dynatrace Intelligence	Product	Deterministic + agentic AI positioning.
S13	Dynatrace agentic AI docs	Product	Permissions, MCP tools, PII handling.
S14	Splunk Observability AI update	Product	Troubleshooting Agent modalities and next steps.
S15	Splunk troubleshooting agent docs	Product	Auto RCA for APM/Kubernetes alerts.
S16	PagerDuty AIOps	Product	Noise reduction, visibility, triage.
S17	PagerDuty Event Intelligence	Product	Up to 98% noise filtering, related incidents, changes.
S18	PagerDuty Spring 2026 release	Product	SRE Agent positioning.
S19	Grafana Assistant Investigations	Product	3.5x faster internal case, agent swarm, evidence trail.
S20	OpenTelemetry AI agent observability	Standard	Agent traces, metrics, logs, GenAI conventions.
S21	AIOpsLab	Benchmark	Interactive microservice/fault/agent evaluation framework.
S22	RCAEval	Benchmark	9 datasets, 735 cases, 15 baselines.
S23	OpenRCA	Benchmark	335 cases, 68 GB telemetry, 3 systems.
S24	CCF AIOPS 2025 RCA paper	Benchmark	HipsterShop-based metrics/logs/traces RCA dataset.
S25	AIOps LLM survey	Research	Taxonomy and literature map.
S26	Auditable graph-guided RCA	Research	Graph/tool-guided auditable RCA.
S27	AIOps “AI Oops”	Risk	LLM-driven AIOps trust/security risks.
S28	TrioXpert	Research	Automated incident management lifecycle.
S29	ZeptoDB repo docs	Internal	Edge profiles, ROS 2, agent memory, time-series architecture.
S30	TechRadar OpenClaw agentic AI security	Risk	Shadow AI, agent identity, permissions, audit concerns.
S31	ITPro / Forrester agentic AI operationalization	Market/Risk	Enterprises remain pilot-heavy due to governance and infrastructure gaps.
S32	TechRadar agentic SOC	Market/Risk	SOC needs machine-speed investigation and response.
S33	Elastic agentic AI SOC	Product/Risk	Closed-loop containment and traceable reasoning.
S34	AgentSOC	Research	Explainable, risk-aware, policy-aligned autonomous SOC.
S35	NVIDIA Jetson Thor	Physical AI	2070 FP4 TFLOPS, 128 GB memory, 40-130 W edge robotics compute.
S36	Physical Intelligence MEM	Physical AI	Multi-scale embodied memory for long-horizon robot tasks.
S37	AIOpsLab project site	Benchmark	Interactive autonomous AIOps agent benchmark implementation.
S38	Autonomous Incident Resolution at Hyperscale	Research	Multi-agent autonomous incident resolution with rollback mechanisms.
S39	OpenTelemetry GenAI observability 2026	Standard	Structured capture of LLM messages and tool calls.
S40	Datadog OTel GenAI semantic convention support	Product/Standard	Prompts, model responses, token usage, tool/agent calls, provider metadata.
S41	Reflexion	Agent memory	Episodic reflection over action feedback without weight updates.
S42	Experiential Reflective Learning	Agent memory	Outcome-based reusable heuristics from experience trajectories.
S43	Voyager	Agent memory / embodied agents	Skill library plus environment feedback for lifelong agent learning.