The question nobody can answer about their agent
Ask a team running a long-lived AI agent a simple question: what does it remember about your business, and when did that change?
Most teams cannot answer it. Agent memory in the current stack is write-mostly and read-opaque. Facts accumulate from sessions, get summarized by the agent itself, and quietly shape every future decision the agent makes. When the agent starts behaving differently in month three, there is no record of which remembered “fact” changed, when it changed, or what changed it.
For a personal assistant, that is a quirk. For an agent that drafts customer messages, schedules work, or feeds an approval queue, it is a governance gap: the agent’s memory is an unaudited input to every governed output.
Memory curation is now native. Governance is not.
The agent platforms are moving fast here. Scheduled memory curation — a background process that reads the memory store plus recent session history and rewrites the memory to merge duplicates, drop stale entries, and consolidate patterns — is now a native capability in the frontier stack. It is genuinely useful: uncurated memory degrades, and an agent reasoning over six months of contradictory notes performs worse than one with a clean store.
But native curation ships as a black box. The curator reads everything, rewrites what it judges stale, and leaves no reviewable record of the edit. From a governance standpoint, that is an unattended process with write access to the input layer of all your other controls.
What governed memory curation looks like
We shipped scheduled memory curation for JieGou managed agents this week — built on the native capability, wrapped in the governance layer it does not ship with:
- Protected namespaces. Portions of the memory store are off-limits to the curator entirely. Operational state, cross-agent shared context, and thread-level records cannot be rewritten by a curation pass, no matter how confident the model is that they are “stale.”
- Hard caps per cycle. A curation pass is bounded: a fixed ceiling on total operations and a lower ceiling on deletions. There is no scenario in which one bad cycle rewrites the agent’s entire worldview overnight. Large changes take many cycles, which means many review opportunities.
- Dry-run previews. A cycle can run in preview mode, producing the full set of proposed changes without applying any of them. New deployments run in dry-run until the operator has seen a few cycles’ worth of judgment.
- Auditable diffs. Every applied change is recorded as a reviewable diff: what was merged, what was replaced, what was deleted, and which sessions informed the change. The question from the top of this post — what does it remember, and when did that change — has a concrete answer at any point in time.
The same release added dual-lane outcome scoring: work produced inside an agent session is scored by the native grader in-session, and drafts produced outside a session go through our judge lane. Every piece of agent work gets an outcome score; memory curation and outcome evidence land in the same audit posture.
Why this matters beyond the feature
There is a pattern worth naming. As agent platforms mature, capabilities that used to be the integrator’s work — memory, scoring, scheduling — become native primitives. That is good. The model layer and its primitives are becoming commodity infrastructure.
What does not become commodity is the governance shell: the scoping, the caps, the approval gates, the replayable record. Native memory curation makes an ungoverned agent better at being ungoverned. The operational question for any team running agents against real customer data is not whether to use the native primitives — use them — but whether every primitive that writes to your agent’s state leaves evidence an auditor, an underwriter, or your own future self can review.
Memory that changes only through capped, diffed, replayable operations is our answer. If your current agent stack cannot show you a diff of what it remembered last month versus today, that is worth fixing before the memory gets interesting.