Governed Enterprise Agent Learning

Enterprise agents learn safely when feedback, outcomes and candidate knowledge pass through provenance, permission, evaluation, approval, staged deployment and rollback controls. Learning is not an agent directly rewriting shared organisational truth because one interaction appeared successful [src-255, src-259].

Six different mechanisms

Mechanism	What changes	Durability	Primary risk
Prompt context	Information supplied to one invocation	Ephemeral	Context injection or leakage
RAG	Evidence retrieved at read time	Source-dependent	Stale, irrelevant or unauthorised retrieval
Persistent memory	Stored state affecting future behaviour	Cross-session	Poisoning, contradiction and permission leakage
Feedback learning	Candidate prompt, workflow, policy or retrieval changes	After promotion	Optimising for noisy or manipulated signals
Fine-tuning	Model weights	Release-specific	Dataset provenance, regression and rollback
Uncontrolled self-modification	Prompts, policies, memory or code without gates	Unbounded	Loss of authority, auditability and stability

The promotion lifecycle

1. Capture feedback, outcomes, incidents and candidate knowledge with provenance.

2. Classify sensitivity, ownership and permitted uses.

3. Deduplicate, reconcile contradictions and check freshness.

4. Evaluate candidate changes offline against representative and adversarial suites.

5. Require approval proportionate to consequence.

6. Publish a versioned prompt, policy, workflow, retrieval rule, memory entry or model update.

7. Deploy through shadow and canary stages.

8. Monitor quality, cost, safety and business impact.

9. Supersede or revoke knowledge when the source changes.

10. Preserve the full decision history and immediate rollback.

Memory is not truth

Useful enterprise memory primitives include:

separate private, team and institutional tiers;
read and write policies evaluated against current identity and purpose;
immutable source and author provenance;
confidence, freshness, expiry and supersession;
candidate and approved states;
retrospective permission checks when access changes;
lineage from derived memories to all supporting evidence ^[src-259].

The Collaborative Memory framework demonstrates several of these primitives, but remains research rather than a broadly validated production standard ^[src-259].

Evaluation gates

No candidate change should be promoted without:

task-success and business-outcome tests;
access-control and provenance tests;
regression and adversarial tests;
cost and latency thresholds;
comparison with the current production version;
human calibration of automated evaluation;
rollback evidence.

Source references

^[src-251] Enterprise managed AI-agent platforms official documentation bundle
^[src-255] NIST AI RMF and EU AI Act controls
^[src-259] Collaborative Memory
^[src-260] Multi-Agent Risks from Advanced AI

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 477 indexed pages and articles.