remember()Keep a fact with provenance.
Memory can carry a subject, source actor, source message, and owning perspective.
Inside the runtime
One message.
An explainable path.
An enterprise agent harness separates model judgment from runtime authority. Follow one team question through identity, source-authorized context, governed tools, and a durable record of the work.
- 01INGRESS
Resolve the room.
Slack or Telegram delivers an event. Mesh normalizes the message and resolves the actor, conversation, thread, and owning agent perspective.
Authorship exists before prompt assembly.actor → conversation → message - 02PARTICIPATION
Decide whether to join.
Direct addressing, thread participation, and conversation shape inform an engage, observe, or ignore decision. Ambiguous name-based addressing can use a bounded classifier.
Observed messages retain the participation decision.engage / observe / ignore - 03CONTEXT
Assemble one agent’s view.
Mesh selects conversation history and eligible memory for this perspective, alongside persona and runtime state. The prompt is a projection of those records, not their source of truth.
Another agent’s private context is not an implicit input.history + memory + persona - 04EXECUTION
Let the model propose. Let the runtime govern.
The model chooses useful work. Mesh validates tool calls, bounds tool rounds, and records model usage and observations. Workspace tools acquire remote, run-scoped resources when needed.
Model judgment and runtime authority have different jobs.model → tools → observations - 05DELIVERY
Return to the conversation.
Output is rendered for the connector and delivered as an attributed response. The run captures its terminal outcome so an operator can distinguish completion, failure, and interruption.
The reply remains connected to the original trigger.reply + usage + outcome
Capture before confidence
The answer is only part of the record.
Inspect the prompt that was assembled, the model calls that followed, their token usage, and the final run state. Debug the path to an answer—not only its wording.
These records exist today. A single, unified operator timeline is planned; this illustration is a reading guide to the underlying data, not a screenshot of a shipped interface.
Where these records liveRUN RECEIPTA reply you can trace.
Illustration- 01TriggerActor and conversation
Ada mentioned Nova
- 02ContextCaptured prompt snapshot
Release thread + eligible memory
- 03ModelRecorded model call
Request, response, token usage
- 04ToolsRecorded tool calls
Arguments and observations
- 05OutcomeDurable terminal state
Completed, failed, or interrupted
Memory is managed state
Remember.
Recall.
Amend.
Forget.
recall()Retrieve what belongs here.
Lexical matching and subject-aware ranking operate within conversation eligibility rules.
amend()Correct without losing the trail.
Versioned state preserves the fact’s revision history when new information replaces it.
forget()Retire what should no longer be recalled.
Forgetting removes a memory from active use. It is not a promise to erase every stored copy or revision.
ROADMAP The next layer
Conversation is a graph.
Complex work should be, too.
The conversation graph records what people said. The proposed WorkGraph describes how an objective executes: bounded tasks, dependencies, verification, and human gates.
WorkGraph, durable commitments, and resumable scheduling are planned capabilities. They are not included as shipped features in today’s beta.
PROPOSED WORKGRAPHNOT YET BUILT
Admitted objective
ResearchImplementCheck
Verify + human gate