Skip to main content

Memory & Knowledge

Your agent becomes more effective over time by remembering what worked in past incidents and referencing your documentation.

SearchMemory queries three sources: past incidents, user memories, and documents to provide grounded responses with citations

How Memory Works

When you ask a question, your agent searches across all knowledge sources simultaneously:

SourceWhat it FindsBest For
Past incidentsSteps that resolved similar issues"How did we fix this before?"
User memoriesFacts you've explicitly saved"Remember my environment uses..."
Knowledge baseYour uploaded runbooks and docs"Follow our standard procedure"

The agent returns a grounded response with clickable citations showing exactly where information came from.


Automatic Learning

Your agent learns from every conversation—no manual training required.

After each session, the agent automatically extracts learnings: symptoms, resolution steps, root cause, and pitfalls to avoid

After each thread completes:

WhatDescription
Symptoms observedError messages, behaviors, patterns
Steps that workedThe resolution path that succeeded
Root causeWhat caused the issue
Pitfalls to avoidWhat didn't work, dead ends

This happens automatically. 30 minutes after a thread goes quiet, the agent evaluates the conversation and indexes the learnings.

Same-Resource Priority

When investigating a resource issue, the agent prioritizes past sessions on the exact same resource:

"App Service app-prod-01 is returning 503 errors"

Your agent first checks: "Have I seen issues on app-prod-01 before?" If yes, those learnings appear first—they have the highest relevance.


Proactive Knowledge Persistence

Beyond learning from completed threads, your agent actively saves what it discovers during conversations. When your agent encounters something important — a tricky configuration, a non-obvious dependency, or a debugging gotcha — it records the insight in persistent knowledge files that carry across sessions.

How it works

Your agent maintains a knowledge directory at memories/synthesizedKnowledge/. A special file, overview.md, is automatically loaded into your agent's system prompt at the start of every conversation. This gives your agent immediate access to the most important context about your environment.

ComponentWhat it does
overview.mdService summary and index — always loaded into context (~2,000 character budget)
Topic filesDetailed notes on specific subjects (e.g., aks-networking-gotchas.md)
Links from overviewoverview.md links to topic files so your agent knows what detailed knowledge exists

What your agent saves

Your agent proactively records insights during conversations:

CategoryExamples
Problem constraints"This service can't scale past 10 replicas due to quota limits"
Strategies that worked"Restarting the pod with --grace-period=0 resolved the stuck deployment"
Strategies that failed"Increasing memory limit didn't help — the issue was CPU throttling"
Non-obvious dependencies"app-frontend depends on a sidecar proxy that must start first"
Configuration details"Production uses custom TLS certificates stored in Key Vault"

Knowledge organization

Your agent organizes knowledge semantically by topic, not chronologically. Each file is a self-contained reference:

FileWhat it captures
overview.mdService summary, key links, index of topic files (~2,000 chars)
team.mdTeam members, roles, expertise (~500 chars)
architecture.mdComponents, connections, environments (~1,500 chars)
logs.mdLog sources, tables, key fields, useful queries (~1,500 chars)
deployment.mdPipeline details, version lookup, rollback procedures (~1,000 chars)
auth.mdAuth mechanisms, identity flows (~800 chars)
debugging.mdCommon issues, troubleshooting guides, runbook links (~1,000 chars)
queries/*.mdExtracted queries organized by topic (~1,000 chars each)

When updating existing knowledge, your agent reads the current file, merges new information, and removes anything that has become outdated or incorrect.

You can ask your agent to save knowledge too

Beyond automatic persistence, you can explicitly ask your agent to save information to its knowledge files:

Save this to your knowledge: our Redis cache uses Premium tier with 6GB,
and failover takes about 90 seconds.

Your agent creates or updates the appropriate knowledge file and links it from overview.md.

This is different from #remember commands (described below), which save discrete facts to a separate memory store. Knowledge files are structured, persistent references that your agent consults at the start of every conversation. User memories are individual facts searchable via #retrieve.


User Memories

Beyond what your agent learns and persists automatically, you can explicitly save discrete facts for your agent to remember. User memories are ideal for environment-specific details that might not come up in incidents but are important for context.

What to save as user memories:

CategoryExamples
Environment facts"Production uses 3 AKS clusters in West US 2"
Team preferences"We prefer CLI over portal for deployments"
Architecture details"app-service-01 depends on sql-prod"
Escalation paths"PagerDuty → Teams channel → phone"

Memory Commands

Manage user memories with these chat commands:

CommandWhat it DoesExample
#rememberSave a fact for future reference#remember our Redis cache uses Premium tier
#retrieveSearch your saved memories#retrieve what's our caching setup?
#forgetRemove a saved memory#forget the outdated Redis info
Example memory workflow

Save important context:

#remember Production uses 3 AKS clusters in West US 2
#remember Our escalation path: PagerDuty → Teams channel → phone
#remember Database failover takes approximately 15 minutes

Retrieve later:

#retrieve how long does database failover take?

Based on saved memory: Database failover takes approximately 15 minutes.


Knowledge Base

Knowledge comes from uploaded documents and MCP connectors, all searchable together

Upload Documents

Go to Builder → Knowledge base to upload your documentation. For full details on supported formats, limits, and how your agent creates knowledge automatically, see Upload Knowledge Documents.

Document TypeGood For
RunbooksStep-by-step incident procedures
Architecture guidesUnderstanding your environment
On-call playbooksEscalation and response procedures
API documentationService-specific knowledge
Team proceduresWorkflow and process docs

Supported formats: Markdown, plain text, PDF, Word, PowerPoint, Excel, images, and more — max 16 MB per file. See Upload Knowledge Documents for the full list.

Connect External Sources

Access knowledge directly from external systems via connectors. See Connect Knowledge for the full list of supported sources.

ConnectorWhat it Provides
Azure DevOpsQuery your ADO wiki pages
GitHubSearch repos, wikis, issues
Custom MCPAny knowledge source you configure

Configure connectors in Builder → Connectors. See Connectors for setup.


Using Knowledge in Conversations

Your agent automatically searches knowledge when relevant:

How should I handle a database failover?

If you have a runbook uploaded:

Based on your Database Runbook (citation link), here are the failover steps:

  1. Verify the health of the secondary replica...

Click citations to view the full source document.


Session Insights

After every thread—whether a sync chat conversation or an async auto-triggered task—your agent generates a session insight. This is how your agent gets smarter over time.

What gets captured

Each session insight extracts structured learnings that become searchable memory:

ComponentWhat it CapturesExample
Symptoms observedError patterns, behaviors"HTTP 503 errors, memory at 95%"
Resolution stepsWhat worked"Scaled up App Service SKU"
Root causeWhy it happened"Memory leak in deployment v2.3"
Pitfalls to avoidWhat didn't work"Restarting didn't help"

When insights are generated

Thread TypeWhenAuto or Manual
Sync chat30 min after last messageAutomatic
Async tasks30 min after completionAutomatic
User feedbackWhen you rate a responseYou trigger it

Viewing session insights

Go to Monitor → Session insights to see:

  • Timeline of agent actions
  • Evaluation scores
  • Key learnings extracted
  • Source thread links — each insight card links back to the thread(s) that generated it, so you can trace any insight to its original conversation

See Monitor Agent Usage for detailed metrics and management.


Best Practices

What to Upload vs Connect

UploadConnect via Connector
Incident runbooksLive wiki pages (ADO, GitHub)
Architecture diagramsSource code repositories
Escalation proceduresReal-time monitoring data
Static API docsFrequently updated docs

Keep Knowledge Current

Outdated documents cause incorrect responses. Review quarterly.

What knowledge documents do you have?

Remove outdated docs in Builder → Knowledge base.

Name Documents Clearly

Don'tDo
doc1.txtproduction-database-failover.md
runbook.mdaks-cluster-scaling-runbook.md
notes.txtescalation-procedures-2026.txt

ResourceWhy it matters
Connect Knowledge →Manage all knowledge sources — files, web pages, and repositories — in one place
Upload Knowledge Documents →Detailed file format support, limits, and agent-generated knowledge
Connect external knowledge sources →How connectors bring external data to your agent
Understand custom agents →How custom agents use knowledge in workflows
Was this page helpful?