Persistence is continuity across work—not just a longer chat
A one-shot assistant receives an input, creates an output, and starts over next time. A persistent agent can carry forward the context, decisions, progress and evidence that matter to later work.
That does not mean keeping everything. It means giving useful continuity a deliberate home, making it correctable, and keeping the result connected to the work that produced it.
Persistent and proactive are different properties
Persistence describes what an agent carries forward. Proactivity describes when it starts work. The two can combine, but neither guarantees the other.
| Reactive: waits for a request | Proactive: starts from a schedule or trigger | |
|---|---|---|
| Non-persistent | A single-session assistant that starts fresh each time. | A stateless alert or scheduled action with no accumulated context. |
| Persistent | An agent that remembers relevant context and continues tracked work when asked. | An agent that uses stored context and a declared trigger within authority and policy. |
Persistence can make proactive work more useful because the agent can reason from what happened before. It does not make proactive action safe by itself: authority, policy and approval still determine whether a consequential effect may happen.
What needs a durable home
| Property | Why it matters | Lumen evidence path |
|---|---|---|
| Context | The agent should not repeat a correction or lose an agreed decision. | Memory and context preview. |
| State | Long work needs an inspectable lifecycle, not an ambiguous spinner. | Run state machine and event log. |
| Evidence | A later reader needs to know what happened and what effect was recorded. | Work Objects, tool calls and receipts. |
| Control | Continuity must not silently broaden permissions. | Approvals and server-enforced authority. |
A persistent agent is useful when work can pause without becoming ambiguous
Consider a research run that compares sources, prepares a recommendation, and proposes an external action. The work is tracked as one run. If it needs a consequential action, the run waits for approval instead of guessing. The request identifies the tool, target and change. When an effect is recorded, the run can keep a receipt with its outcome and evidence reference.
- Instruction
- Plan and gather
- Waiting for approval
- Approved or refused
- Effect and receipt
- Work Object
The result is not “the agent kept working somehow.” It is work with a state, a reason for a pause, an authority boundary and evidence a person can return to.
How Lumen makes persistence inspectable
Lumen keeps workspace context separate from the model that happens to answer. It represents multi-step work as a run with a declared lifecycle and event log, keeps approval decisions with the run, records tool-call receipts, and exposes the producer/checker relationship for verified answers.
The product is explicit about its boundaries: rooms are single-workspace today, and Public API v1 is in development until it is proven end to end.
What persistence does not guarantee
Lumen does not claim Tier B always-on continuity across host sleep, VM power-off, or a laptop closed without an always-on worker. Scheduled and automation work has Tier A process-restart recovery coverage; that is a narrower, documented guarantee.
Persistence is not unlimited authority, hidden context that is automatically safe, external verification of every effect, or independent verification by default. Lumen labels a same-model check as a self-check rather than presenting it as independent review.
Questions to ask before you call an agent persistent
- Where does useful context live, and can a person correct it?
- Can multi-step work show its state, reason for pause and legal transitions?
- Does continuity preserve evidence, or only chat text?
- Does stored context silently broaden what an agent may do?
- What survives a process restart, and what does not survive host sleep or power-off?
- Can a reader distinguish a proposal, approval, effect and receipt?
Common questions
Is a persistent agent the same as an always-on agent?
No. Persistence describes continuity of useful context and work state. Always-on adds an operational claim about hosts and workers that must be verified separately.
Is persistence the same as memory?
No. Memory is one form of continuity. A persistent system also needs lifecycle state, evidence about effects, and controls that keep context from silently becoming authority.
Do persistent agents act on their own?
Not necessarily. Proactivity is a separate property driven by a declared schedule or trigger. Consequential effects still require the authority and approval policy that applies to the run.
What should be kept after a run?
Keep the context, decisions, lifecycle state, outputs and evidence that a person can inspect and correct. Do not treat an unbounded hidden transcript as a system of record.
Inspect the work, not just the answer
Read How Lumen works, inspect Build with Lumen, or review Everything Lumen does, surface by surface.