Agent Infrastructure Notes #3

Your agent needs a memory write policy.

Published July 9, 2026 · For agent builders, platform teams, AI consultants, and ops/security reviewers

Core idea: The dangerous memory bug is not forgetting. It is remembering too easily. Before an agent can turn temporary context into durable operating truth, it needs a memory write policy.

The short version

Memory is not a bigger context window. Memory is authority over future behavior.

If one bad run can write durable facts, preferences, permissions, client constraints, source mappings, or workflow decisions, that run can quietly poison every later run.

Production memory needs explicit states:

Without those states, “the agent remembered it” becomes a security and operations problem.

The unsafe pattern

“Remember anything useful.”

That is convenient during a demo and dangerous in production.

The agent may preserve stale facts, temporary instructions, user frustration, incorrect tool outputs, hallucinated conclusions, or one-off exceptions as durable policy. Worse, it may not record where the memory came from or what it replaced.

“You may propose memory updates. Durable writes require source, scope, owner, expiry, conflict check, and approval path.”

That is the control boundary. The goal is not to prevent memory. The goal is to make memory accountable.

A memory write is an external action

Teams often treat memory as internal model state. I think that is wrong. A durable memory write changes how the system will act later. It deserves the same seriousness as writing to a CRM, sending a message, or changing a config file.

Memory questionWhy it matters
What changed?Lets a reviewer inspect the delta, not just the final memory.
What source authorized it?Prevents rumors, drafts, and hallucinations from becoming policy.
What scope does it apply to?Stops one workflow's rule from leaking into another.
What did it replace?Prevents silent overwrite of older operating truth.
When does it expire?Stops stale facts from driving future actions.
Who owns review?Creates accountability when the memory is wrong.

The minimum viable memory write policy

  1. Allowed memory classes — identity, preferences, workflow facts, source maps, permissions, lessons, and non-goals.
  2. Forbidden memory classes — secrets, credentials, private third-party data, speculative conclusions, one-time instructions, and unverified facts.
  3. Promotion path — observed → proposed → approved → durable, with explicit rules for skipping stages.
  4. Provenance — source URL/file/message/task, timestamp, run ID, and confidence.
  5. Scope — user, project, workflow, client, tool, or session boundary.
  6. Conflict handling — what happens when new memory disagrees with old memory.
  7. Expiry and review — time-to-live, stale markers, owner, and renewal criteria.
  8. Rollback — how to revoke or restore previous memory if a write was wrong.
  9. Receipts — every durable write should appear in the run receipt.

The sellable buyer wedge

Memory governance is a strong first audit because it exposes whether a team is ready for autonomy before the agent gets dangerous tools.

Ask one workflow:

If those answers are unclear, the system is not ready for broader delegation yet.

Free artifacts

If you already have a specific workflow whose memory layer needs review, the next step is the AI Agent Memory Audit or the Custom Agent Audit Mini-Report.

My operating rule

If a memory can change future behavior, it needs provenance, scope, expiry, and a rollback path.