Skip to main content

Memories

A memory is one fact or lesson a past run established, with a note on when it applies. When a later run hits a similar situation, the agent recalls the memory instead of working it out again. An incident your team has seen before gets resolved from what the last run found, with fewer tool calls spent rediscovering your environment.

Memories are part of the Engineering Context Graph, alongside skills and the entities. You don't write memories. Learn generates them from finished runs, and you review them here.

Where memories come from​

Learn reflects on each finished run and proposes a memory when the run established something the next run should know:

  • An accepted root cause. A diagnosis a person confirmed becomes knowledge for the next similar incident.
  • Your guidance. "Ignore the database, check the deploy" is carried forward to the next run that faces the same choice.
  • A hard-won discovery. A non-obvious cause, or a fact about your environment that took many tool calls to find.

A proposal can also replace an outdated memory, or archive one the run proved wrong. Most routine runs propose nothing.

Lifecycle​

Every memory has one of five statuses.

StatusMeaning
PendingWaiting for review. Not recalled by any run yet.
AppliedLive. Recalled by future runs that its owner allows. You approved it, or it was auto-accepted.
DismissedYou declined it, optionally with a reason. Never recalled.
Consolidated into skillYou folded it into a skill. It's still recalled alongside applied memories.
SupersededA newer memory replaced it during deduplication. Archived automatically.

The Memories page, in Context Graph, shows the number of pending memories. That queue is the only part that needs your attention.

One agent's memories​

An agent's page in Agent Studio has a Memories tab. It lists the memories that agent's runs proposed, in the same views as the Memories page: Need your Review, Active, Dismissed and Archived.

On the Memories page, the Agent filter narrows the lists to one agent. Memories from investigations don't belong to an Agent Studio agent, so they're hidden while the filter is on, along with the other filters and search, which only apply to them.

When a memory skips review​

A pending memory is applied without review when one of these is true:

ReasonCondition
Supported by multiple investigationsThe same finding came up in enough near-duplicate investigations to clear your support threshold
Human guidanceIt came from direction a person gave during a run
Accepted hypothesisIt came from a hypothesis a person accepted
Debugging heuristicIt's a general debugging technique, not a fact about one incident
Integration relatedIt's about a connected integration

You can turn each reason off, and you set the support threshold.

Some memories are held for review even when one of the weaker reasons applies: ones the model judges obvious, specific to one incident, or textbook. Support from multiple investigations still applies them, because repeated findings are evidence. And if a person explicitly asks the agent to remember something, that memory is never held or dropped on those grounds.

From memories to skills​

A memory records what one run found. When several runs repeat the same diagnostic work, write it once as a skill instead. See Memory or skill? for how to choose.

You can fold an applied memory into a skill yourself, which sets it to Consolidated into skill. Learn will also generate skills from runs you pick.

Skills are never auto-accepted. An agent follows a skill step by step, so every new or changed skill waits for an engineer's review.

  • Learn: how memories are generated from a run
  • Skills: procedures you write or approve
  • Benchmark: why benchmark runs won't create memories
  • Capabilities: which agents can recall and write memories