Skip to main content
A Trust Posture is Mnemom’s fleet-level oversight policy artifact. It’s the named object that decides which sideband detectors fire, at what thresholds, on what cadence, with what severity, with what fan-out. Postures are assigned to teams; the observer reads each team’s effective posture at sweep time and runs the corresponding detectors against the team’s members. If you’ve used CrowdStrike Falcon’s Prevention Policies, CloudFlare’s Rulesets, Okta’s Policies, Microsoft Defender’s Configuration Profiles, ServiceNow’s Domains+Policies, or Zscaler’s Policy Sets — Trust Posture follows the same shape. A posture is a first-class, library-cataloged, versioned object; you assign it to teams; you clone it to customize; you roll back when needed.

​
The named-object pattern

Three properties matter:
  1. Posture is the entity, not a property on the team. The posture lives in a library. Teams are assigned a posture. The posture outlives the team↔posture relationship — clone it, edit it, A/B-test it, export it, archive it, all without touching any team.
  2. Every mutation creates a new revision. Postures are versioned. The current revision is what observer reads; older revisions stay queryable. Rollback is forward-only — rolling back to revision N creates a new revision whose body equals revision N’s. Audit linearity preserved.
  3. Mnemom ships opinionated defaults. Three platform-scope postures arrive seeded:
    • Standard — sensible enterprise defaults (the right starting point).
    • High-Compliance — banking-core / regulated industry shape; tighter, faster, louder.
    • Low-Latency — real-time agent UX; permissive thresholds, slower cadence.
    Customers either use these as-is or clone-and-customize. There’s no “empty config” starting state.

​
Why postures are separate from cards

Trust Posture vs. Cards covers this in depth. The short version:
  • Cards (Alignment Card, Protection Card) are agent-scoped runtime treatment. They answer “what is this agent allowed to do?” and “how do we screen this agent’s content for threats?” Read by the gateway at request time.
  • Postures are team-scoped policy input. They answer “across this group of agents, what fleet-level patterns warrant alarm?” Read by the observer at sweep time.
The two artifact types are parallel — they share infrastructure (composer, KV cache, audit log) but their data planes are independent. A posture never folds into a card; a card never folds into a posture.

​
The cascade

Effective posture is composed Platform → Org → Team:
LayerSourceNotes
PlatformOne of tp-platform-{standard,high-compliance,low-latency}Cascade-closure floor. Every team has at least the platform layer.
OrgThe org’s chosen default (set on orgs.default_posture_id)Optional. NULL ⇒ org doesn’t override the platform layer.
TeamThe team’s directly-assigned postureOptional. NULL ⇒ team inherits org/platform.
Strictest-wins per axis:
  • enabled (bool): OR-true wins (any layer enabling fires the detector)
  • cadence_seconds (number): min wins (shorter = more frequent = stricter)
  • fire_on.* (number or null): min-among-defined wins; null contributes nothing
  • severity_floor (enum): min wins (lower floor = more advisories)
  • severity_on_fire (enum): max wins (higher = louder advisory)
  • Pattern booleans (patterns.outliers, cluster_partition): OR-true wins
Cascade closure for posture guarantees every team has an effective posture; the platform default is the floor.

​
Library, editor, and assignment

Three first-class management surfaces:
  • Library (/settings/organization/postures): browse the three Mnemom defaults plus your org’s customized postures. Cards display assigned-team count, current revision, last-edited-by, divergence-from-default indicator.
  • Editor (/settings/organization/postures/{id}): three-column composition preview. Your draft, the org floor, the effective composed result. Per-field tightening callouts.
  • Assignment (on the team detail page): a tree view of Platform → Org → Team showing which posture is at each layer and which axis comes from where.
CLI mirror via the mnemom posture subcommands (list, show, create, update, clone, revisions, diff, assign, unassign, preview-compose, delete).

​
What postures don’t do

Postures are deliberately not a place to set per-agent runtime treatment. If you want to:
  • Forbid rm -rf on a specific agent → that’s Alignment Card enforcement.forbidden_tools.
  • Block PII from leaving via an agent’s responses → that’s Protection Card DLP thresholds.
  • Cap autonomous spend → that’s Alignment Card autonomy.max_autonomous_value.
If you want to:
  • Detect when a team’s value coherence degrades → posture sideband.coherence.
  • Detect fault lines forming between agents → posture sideband.fault_line.
  • Detect fleet outliers — agents drifting from the group’s center of mass → posture sideband.fleet.

​
Audit + compliance

Every posture mutation emits a governance_audit_log row with target_type='posture' and one of:
  • posture.create, posture.put, posture.clone, posture.delete
  • team_posture.assign, team_posture.unassign
Versioning + audit + assignment metadata together give compliance reporting a queryable foundation: “What was our coherence detection threshold for the trading-desk team on 2026-03-31?” is one SQL join.

​
Further reading