Skip to content

Agents

Status: Designed, not yet implemented.

In the memory architecture, agents are the workers responsible for maintaining the store. The infrastructure layer provides content-addressing, atomic writes, search, and the event log. Higher-level operations such as synthesis, decay, and conflict resolution are performed by those workers according to policy documents.

Each agent has a name (unique within the colony), a role (defines capabilities and authority), and an author identity (every write is attributed). Multiple agents work concurrently.

A minimal deployment has five roles. The names below are mnemonics, not required implementation terms:

RoleResponsibilityTrigger
ScribeIngest and parse raw input into typed objectsNew data arrives
AnalystSynthesize patterns from accumulated raw materialObject count exceeds threshold
ArchivistDecrement confidence scores, archive stale objectsPeriodic schedule
EditorDetect and resolve contradictionsConflicting objects surface
StewardMonitor health metrics, propose convention changesMetrics degrade

New roles are added by writing a policy document that defines the role’s triggers, procedures, and authority.

An agent operates by:

  1. Reading applicable policies from the store
  2. Checking if trigger conditions are met
  3. Executing the referenced procedure
  4. Writing the result with provenance (author, source objects, confidence)
  5. Appending the event to the log

The policy defines the trigger, the procedure defines the work, and the output schema defines the result. In this design, changing the documents changes the behavior.

When the system encounters gaps — unanswered queries cluster in a domain, or health metrics degrade in a specific area — new agents can be created:

TypeTriggerMechanism
RenewalContext exhaustionFresh agent takes over with compressed context from predecessor
SpecializationDomain mismatchGeneralist spawns a domain specialist
DelegationRecurring subtaskRepeated work pattern justifies a dedicated agent

System size, inference budget, and spawn approval are human-configurable parameters.

Open questions: At what scale do coordination costs outweigh specialization benefits? How should the colony balance inference budget across agents? These are empirical questions the implementation needs to answer.