Custom Conscience Values: Per-Org Alignment Policies for Enterprise

Mnemom Research | February 2026

Every industry has alignment constraints that generic defaults can't capture. A healthcare agent needs to know that patient safety outweighs efficiency. A fintech agent needs to understand that regulatory risk is a hard boundary, not a tradeoff. A legal agent needs to treat privilege preservation as inviolable.

Today we ship Custom Conscience Values -- per-org alignment policies that inject your domain-specific requirements into every AIP integrity check. Define what matters most for your organization, and the conscience layer evaluates every agent interaction against those values automatically.

This is the second enterprise capability after the Agent Containment Engine. Containment answers "what do I do when an agent misbehaves?" Custom conscience values answer "what does misbehavior mean for my organization?"

What Are Conscience Values?

AIP integrity checks already evaluate agent behavior against a set of conscience values -- ethical guidelines that the analysis model uses to assess reasoning quality. Until now, these were system defaults: general-purpose values covering honesty, harm avoidance, and autonomy boundaries.

Custom conscience values let you add your own. Each value has:

Name: A short identifier (e.g., "Patient Safety First")
Description: What this value means for agent behavior (up to 500 characters)
Type: BOUNDARY, FEAR, COMMITMENT, BELIEF, or HOPE
Severity: Advisory (flagged in analysis) or Mandatory (triggers boundary violation)

The five types map to different roles in conscience evaluation:

Type	Role	Example
BOUNDARY	Hard limits that must never be crossed	"Never recommend actions that could harm patients"
FEAR	Risks the agent should be wary of	"Regulatory penalties from non-compliant advice"
COMMITMENT	Promises the agent upholds	"Always disclose when providing medical information"
BELIEF	Core principles guiding behavior	"Patient autonomy and informed consent are paramount"
HOPE	Aspirational outcomes to work toward	"Improved health outcomes through accurate information"

Layered Resolution

Values are resolved in three layers, and the layering is deliberate.

Layer 1: Base. The system default conscience values. These cover universal alignment concerns -- honesty, harm avoidance, respecting autonomy. They're the foundation.

Layer 2: Org. Your custom values. These are always applied to every agent in your organization. Org policy is mandatory -- individual agents cannot opt out. This is where you encode domain-specific alignment requirements.

Layer 3: Agent. Per-agent values from alignment cards. These are additive on top of the org layer. An individual agent might have additional specificity ("this agent handles pediatric cases" on top of the org-wide "patient safety first"), but it cannot remove org policy.

Augment vs. Replace

The org layer supports two modes that control how your values interact with the system defaults:

Augment (default): Your org values are added alongside the system defaults. The agent is evaluated against both. This is the right choice for most organizations -- you're adding domain-specific requirements without losing the general-purpose safety net.

Replace: Your org values replace the system defaults entirely. The agent is only evaluated against your custom values (plus any per-agent values). This is for organizations that need complete control over the conscience prompt -- typically those with mature internal alignment frameworks that already cover the ground the defaults address.

Mode only controls org values vs. defaults. Per-agent values are always additive on top.

The API

Seven endpoints under each organization's route:

GET    /v1/orgs/:org_id/conscience-values          # List all values
POST   /v1/orgs/:org_id/conscience-values          # Create a value
PATCH  /v1/orgs/:org_id/conscience-values/:id      # Update a value
DELETE /v1/orgs/:org_id/conscience-values/:id      # Delete a value
PUT    /v1/orgs/:org_id/conscience-values/order    # Reorder values
PUT    /v1/orgs/:org_id/conscience-values/mode     # Set augment/replace + enable/disable
GET    /v1/orgs/:org_id/conscience-values/log      # Audit trail

Create a value:

{
  "name": "No Regulatory Risk",
  "description": "Never recommend actions that could expose the organization to regulatory penalties, sanctions, or compliance violations.",
  "type": "BOUNDARY",
  "severity": "mandatory"
}

The mode endpoint controls both the augment/replace behavior and the global enable toggle:

{
  "mode": "augment",
  "enabled": true
}

Limits

Maximum 20 values per org
Name: 1-50 characters
Description: 1-500 characters
RBAC: Owner and Admin can create/update/delete. All roles can view.

How It Flows

When an agent makes a request through the gateway or observer, the integrity check resolves conscience values through a single RPC call (KV-cached at 5 minutes, fail-open):

Look up the agent's org
Check if custom conscience values are enabled and the plan includes the feature
Fetch active values ordered by sort order
Apply layered resolution: base + org + agent
Pass the merged values to the AIP analysis

The same resolution runs in all three paths: the gateway (real-time checking), the observer (post-hoc analysis), and the analyze API (hybrid mode). Your org's alignment policies are consistently enforced regardless of how the integrity check is triggered.

Audit Trail

Every mutation to conscience values is logged:

Actions: create, update, delete, activate, deactivate, reorder, mode_change
Actor: Who made the change
Changes: Old and new values for every modified field
Timestamp: When it happened

The audit log is queryable through the API and visible in the org settings dashboard. For SIEM integration, every mutation also fires a conscience.values_updated webhook event.

The Dashboard

The org settings page now includes a Conscience Values section. From here you can:

Toggle custom values on/off
Switch between augment and replace modes
Create, edit, and delete values with type and severity controls
Toggle individual values active/inactive without deleting them
View the full audit log in a side panel

Examples

Healthcare

[
  {
    "name": "Patient Safety First",
    "type": "BOUNDARY",
    "severity": "mandatory",
    "description": "Patient safety must always take priority over efficiency, cost reduction, or throughput optimization."
  },
  {
    "name": "Clinical Evidence Required",
    "type": "COMMITMENT",
    "severity": "mandatory",
    "description": "All medical recommendations must be grounded in peer-reviewed clinical evidence."
  },
  {
    "name": "HIPAA Compliance",
    "type": "BOUNDARY",
    "severity": "mandatory",
    "description": "Never process, store, or transmit PHI outside of approved systems and workflows."
  }
]

Fintech

[
  {
    "name": "No Regulatory Risk",
    "type": "BOUNDARY",
    "severity": "mandatory",
    "description": "Never recommend actions that could expose the organization to regulatory penalties."
  },
  {
    "name": "Fiduciary Duty",
    "type": "COMMITMENT",
    "severity": "mandatory",
    "description": "Always act in the best interest of the client. Prioritize client outcomes over internal metrics."
  },
  {
    "name": "Audit Trail Integrity",
    "type": "BOUNDARY",
    "severity": "mandatory",
    "description": "All financial decisions must be traceable and auditable. Never circumvent logging."
  }
]

Availability

Custom conscience values are available on the Enterprise plan. The feature is gated via the custom_conscience_values flag -- non-enterprise orgs receive a 403 feature_gated response with an upgrade link.

This is an Enterprise-only feature because org-wide alignment policy enforcement is inherently a governance concern. Teams evaluating Mnemom for regulated industries -- healthcare, finance, legal, government -- need this level of control. It's also a prerequisite for the Governance-as-Code (E-12) and SOX/HIPAA compliance bundles (E-13) shipping later this quarter.

Get Started

If you're on an Enterprise plan, the API is live now. No migration needed for existing agents -- custom values are additive by default.

Docs: Custom Conscience Values guide -- full explanation with examples
API reference: docs.mnemom.ai -- 7 endpoint definitions with request/response schemas
Dashboard: mnemom.ai/dashboard -- manage values from org settings
Changelog: docs.mnemom.ai/changelog -- technical release details

Default conscience values work for general-purpose agents. But when your agents operate in domains where the wrong answer has regulatory, medical, or financial consequences, generic isn't good enough. Now you can tell the conscience layer exactly what matters.

Mnemom builds alignment and integrity infrastructure for autonomous agents. Custom conscience values are part of the mnemom gateway, available on Enterprise plans.

GitHub: github.com/mnemom