Do AI agents need a second brain? Start with what must survive
Contents
Does your agent need memory?
Start with what must survive
An agent needs durable state when something useful must survive a run. It does not automatically need a new memory service, and restricted data is a separate access question.
Ask what would be lost when the run ends
An agent reads a permitted invoice record, formats a summary and returns it to the requester. It handles restricted information, so access controls matter. But if the next task can read the current source again, it may have no reason to save a new memory of this one.
Now change the task: the agent must pause while a person supplies a missing purchase-order reference and resume the same review tomorrow. The unfinished work needs durable state. That still does not mean it needs a general-purpose knowledge graph or automatic transcript capture.
Five persistence decisions
These constructed cases illustrate a decision rule, not measured product results.
| Task | What must survive | Smallest plausible approach |
|---|---|---|
| Summarize one current permitted record | Nothing beyond the result and any required operational record. | Read the source for each task; no new semantic memory required. |
| Follow repository build conventions each session | Stable team instructions. | Maintain reviewed instructions in the existing repository. |
| Resume an incomplete approval workflow | Task identifier, current stage, relevant references and outstanding decision. | Persist a workflow checkpoint or task record. |
| Remember a user-approved formatting preference | A preference that can be inspected, corrected and removed. | An explicit preference record may suffice. |
| Answer what the team decided across tools | Durable decision evidence and a retrieval path. | Start with existing decision documents; add another store only for a demonstrated gap. |
An existing file is already persistence
A note saved between runs is not “starting from zero.” A repository’s instructions, a task system and a source document may already carry the continuity an agent needs. The problem can be failing to load the right material rather than lacking a new database.
Claude Code, for example, documents persistent instruction files and auto memory; project instructions can be shared through source control. That is enough to disprove the general claim that persistence must live in an external service. Claude Code memory documentation
For a coding project, first record the real build convention in the existing team instructions and verify that the next session loads it. Add an external store only if you need something the current arrangement does not provide, such as a supported retrieval path across several clients.
A pause-and-resume example
Imagine a fictional agent reviewing request R18. It has checked two required attachments but is waiting for a reviewer’s decision. Its durable record says: R18; stage = awaiting review; attachments = A2 and A3; decision = not received. That record lets tomorrow’s run resume without treating the missing decision as approval.
It does not need to preserve every exploratory sentence the agent generated. When the reviewer responds, the workflow validates and records the decision, then advances. The saved record should identify current task state rather than turning a draft inference into a permanent fact.
If two agents can update R18, coordination becomes an additional requirement. The shared-memory version-conflict example shows why shared storage alone does not prevent stale writes.
Decide access independently
A stateless agent can retrieve sensitive records under per-request authorization. A persistent memory service can contain only public facts. Persistence and sensitivity are separate dimensions; neither determines the other.
Adding memory creates another place to retain, correct and potentially expose information. Choose what to store deliberately. Do not enable indiscriminate capture just because continuity sounds useful. The accountable-memory guide explains why deleting one stored record does not prove every derived copy is gone.
Finish the design decision with one sentence: “This fact or task state must survive because the next run cannot reliably obtain it from the existing source.” If you cannot complete that sentence, improve retrieval or task instructions before introducing another memory layer.
About this article
Written by the HeyBrain Editorial Team with AI assistance. The examples are fictional and illustrate design choices; they are not customer stories, product demonstrations or measured test results. Primary sources are linked beside the claims they support.
