doublegate / Agent memory
A company’s agents relearn the same lessons. Keep the ones a reviewer checked.
doublegate holds what an agent writes until a person or model other than its author signs it. Other agents can then find it if they search every work area. By default an agent searches its own area and those above it.
Available today: the gate, its console, a Claude Code plugin and the offline SDK.
Each point is a claim, one thing written down for reuse. Select one to see what happened and why.
The map follows one worked example: a team's alert policy and what happened to each revision.
Selected claim
every alert has an owner team, a runbook link and a severity; the linter rejects an alert missing any of them.
New — promoted in the payments sandbox work area, because the policy it cites requires all three fields.
Revision 4 is held and reviewed as a claim of its own. Follow revision 4.
Claims shown so far: 0.
The next session pays to learn it again, in model usage and in someone's time. A six-month field experiment across 66 firms found AI tools saved individuals time. The researchers detected no shift in the amount or mix of their tasks (AER: Insights). Sharing what agents learn is a separate change, and its effect on a business is unmeasured.
The example is one team's alert policy for its payments sandbox. Revision 3, the source of the map's selected claim, requires an owner team, a runbook link and a severity on every alert.
Revision 4 adds a rollback test and is held. The grading model's two readings disagree, so it waits for a person, who rejects it: the test it cites is missing. Revision 3 stays promoted. A promotion never carries over to the next revision.
Nothing is shared knowledge until a reviewer other than its author has signed it.
All five run on one gate, the doublegate service on the operator’s machine, and each has a cost.
A Claude Code session writes what it learns into its own work area. A new session finds an earlier session's promoted claims when it recalls across the space. Recall runs when the model calls it. Recall before every prompt is on the roadmap.
The Claude Code pluginA person writes a claim in the gate's web console, on the gate's machine. It is reviewed like an agent's, and its author can never sign it.
Review in the consoleA skill is instructions, sometimes with scripts, that change what an agent does. A skill file in a folder the operator keeps is sent for review by the skill watcher and can be promoted here. A bundle imported as an archive is inspected without running anything. It stays held on this gate, counting toward its limit of 10,000 held claims.
Why skills are reviewed before reuseThe gate never counts a signature from the identity that wrote a claim: an agent, a person or the operator. The console shows its author Refused: the claim's author cannot sign it
in place of Sign. An agent-role API key reaches 37 of 83 published methods. None of them promotes, signs, rejects, demotes or hides, and no method deletes. If an agent starts as the gate's own operating-system user, its shell can promote what its session wrote (how to prevent that). Run the agent as another operating-system user and the rule holds against it. A gate that runs under a service user of its own is on the roadmap.
Recall returns promoted claims and the asker's own held ones, marked as held, never another author's held claim. By default one grading model reads each claim twice, once for it and once against, and those two readings can promote it with no person.
How to check itThe journal records every claim written, promoted, rejected or demoted as a signed event. Each entry links to the one before, so a later change shows. One command checks every signature. Nothing is deleted, so the record only grows.
How to check itWithout the reviewer role, a caller asking for another author's held claim, by id, listing or count, is answered as if it did not exist. We publish the four things it can still learn. Three are that the quarantine is full, that it is more than 1,000 events behind, and roughly how long the journal is. The fourth takes many timed calls: whether a claim exists whose id it holds or can compute. An id is a digest of the whole write. What a caller can still learn, measured →
How to check itPromotion decides whether recall can return a claim, and adds no weight to its order. Neither does the provisional mark. By default recall matches words. Matching by meaning needs an embedding model the operator configures.
How to check itEvery external number links its primary source on the evidence page.
Every agent on a gate shares what it holds, and every promotion is that gate's own, marked provisional. Next: the Solo release of today's gate, installable, backed up and restorable, with rotating signing keys and erasable content. Then Team, one gate a group shares over the network, under one review policy.