Skip to main content
Mnemom’s alignment and protection cards are composed of independent primitives — values, modes, principal, autonomy, capabilities, conscience, enforcement, audit on the alignment side; mode, thresholds, screen_surfaces, trusted_sources on the protection side. Each primitive has its own write surface:
You can update one slot of a spec without rewriting the whole card. The composer recomposes after every write; downstream agents pick up the new effective state on their next request.

​
The shape

The full set is 12 primitives × 4 scopes × 2 verbs = 96 endpoints. Every endpoint shares the same contract:

​
PUT vs PATCH

PUT replaces the entire primitive. The body is the new full value of that spec key.
PATCH applies RFC 7396 JSON Merge Patch. null deletes a key, missing keeps it, present replaces it. Arrays are replaced wholesale (RFC 7396 §1, not item-merged).
The PATCH above adds escalation_contact to the existing principal block. The type, identifier, and relationship fields are kept; the response carries the merged shape.

​
The primitive list

​
Alignment (8 primitives)

​
Protection (4 primitives)

​
Recompose-on-write

Every sub-resource write triggers a recompose so the agent’s effective card reflects the change immediately.
  • Agent scope — recompose runs synchronously in the request lifecycle. The response carries the new content_hash + version.
  • Org / team / platform scope — fan-out: every member agent is marked for recompose. The response carries agents_flagged_for_recompose: N. Background workers drain the queue over the next few seconds.

​
Splice-and-validate-whole

Writes don’t blindly accept your patch. The composer reads the current full spec, splices your primitive in, then validates the merged result against the unified card schema.
  • Errors inside the primitive you wrote — 400 with a care-framed message. The write doesn’t land.
  • Errors outside the primitive (pre-existing structural problems in unrelated keys) — surface as _warnings on the 200 response. The write lands; you’re alerted to issues elsewhere in the spec that would benefit from a follow-up.
This means a PATCH on principal is never blocked by a validation problem in, say, audit — but you’re told about the unrelated problem so it can be addressed on its own write.

​
Care-framed errors

Every error path returns a structured JSON body with care-framed prose. The runtime asserts no compliance language slips through; the test suite scans the source modules for the forbidden words.
The vocabulary on every customer-facing message: would benefit from, depends on, run X then retry, you’d be supported by.

​
When to use sub-resource verbs

  • Operator updates one slot — bump retention_days, change the principal’s escalation_contact, add a forbidden tool. Sub-resource verbs are surgical.
  • CI Action edits a single primitive — the mnemom/cards-action drives sub-resource verbs from per-file YAML diffs.
  • Programmatic auditing tooling — a script that audits a fleet and tightens one mode value across many agents.
For the full-spec replace path (publishing a new manifest end-to-end), use the root-level PUT /v1/<resource>/<scope>/<scope_id> documented in Cards as Resources. For previewing a write before committing, use the simulate endpoint.