Skip to main content

Context and memory

Consuelo OS gives agents the context they need to do useful work without loading the entire workspace into the conversation. Context is what the agent needs for the current task. Memory is reusable context that helps the agent behave better over time. Structured records remain the source of truth.

Use structured records first

Structured records should be used when exactness matters. Keep these structured:
  • people
  • companies
  • calls
  • notes
  • tasks
  • opportunities
  • appointments
  • artifact records
  • approval records
  • execution logs
  • integration state
  • workspace identity
These records should be read through approved API facades when OS is connected to the app.

Use memory for supporting context

Memory is useful for context that helps the agent work better but does not need to be the canonical business record. Good memory examples include writing preferences, customer language patterns, prior strategic decisions, reusable summaries, team preferences, playbook notes, and known workflow conventions. Memory should not be the only copy of business-critical records.

Context sources

SourceUse forNotes
Structured API dataExact records and relationshipsSource of truth for workspace objects.
Artifact metadataGenerated outputs and provenanceLinks results back to executions.
Local filesLocal-only workflows and self-hosted useShould be selected or scoped.
MemoryReusable context and preferencesSupport layer, not authority.
Vector contextFinding relevant textUseful for retrieval, not final truth.
Capability healthWhat OS can currently usePrevents skills from failing late.

What context should include

For most agent sessions, OS should provide workspace identity, runtime mode, available skills, capability health, artifact storage mode, approval behavior, useful docs links, and recent execution summary when useful. The goal is to give the agent enough context to work correctly, not to preload every record.

Design rule

Structured first. Memory second. Vector context third. If a value needs to be exact, query the structured source. If a value helps with judgment or recall, memory can help. If a value needs broad text retrieval, vector context can help locate it before the skill checks structured sources.