ADR 3: No cross-session memory

Accepted, 2026-08-25.

Context

Memory was the plan’s third pillar: learned app layouts and successful traces, so an agent that has reached Wi-Fi settings once does not rediscover the route.

The S2 numbers narrowed it before anything was built. The agent sits at a hand-written oracle’s action floor on eight of ten tasks, so a remembered route recovers nothing. The entire remaining gap across the set is five actions, all of them on the dead-switch injection, where roughly six actions go on discovering that a control accepts input and does nothing.

So the implementation remembered failures rather than routes, which was the only version the numbers could support. Notes came from the verifier rather than from the model, a single no-op was never recorded, and a control that moved deleted its note immediately.

What was measured

Nothing in the suite measured a repeat, because every task starts from a fresh device and a fresh agent. A harness was added to run a task twice with notes carried across, three trials per arm, on gpt-5.6-sol. Actions on the second encounter:

framing	second encounter	outcome
no memory (control)	7 [6, 13, 7]	passes
hedged	11 [11, 16, 7]	passes, worse than no memory
assertive	3 [6, 3, 0]	fails

enable_bold_text, the control task with no headroom, was 3 actions in every run of every arm. Memory neither helped nor hurt it.

The hedged briefing said a control “may not work”, told the agent the screen outranks the note, and to ignore a note that turns out to be wrong. It measured worse than remembering nothing. The reading that fits: a hedged note does not remove the investigation, it motivates one. The agent is handed a suspicion and spends actions confirming it.

The assertive briefing said to treat the notes as settled and not re-test them. It was much faster, and one run in three took zero actions: the agent read the note, concluded the switch was dead, and finished without touching the device. Reproduced twice more directly, deterministically.

That run failed the suite, and it failed on a guard added in S2 for an unrelated reason: an unachievable task only passes if the agent actually attempted it. Without that guard this would have been recorded as memory making the task three times cheaper.

Decision

No cross-session memory. memory.py and its wiring are removed.

The two framings are not two attempts at one design, they are the two ends of one dial, and the measurement says the dial has no good setting:

• Hedge it enough to be safe, and it costs more than not remembering.
• Assert it enough to save the work, and the agent stops checking the device.

The second is the serious one. A control noted as dead may have been fixed, the app may have been updated, or the note may have been wrong when written. An agent that skips the attempt reports a failure it did not observe. On a simulator that is a bad number; on someone’s phone it is a confident false statement about their device.

That is the same principle the perception layer already enforces, where RefTable exists so a stale ref is detected rather than trusted, and every action re-reads the screen before acting rather than reusing coordinates from the last observation. Remembered state that outranks live perception is the thing this codebase has spent the most effort avoiding, and cross-session memory is that idea with a longer time horizon.

Alternatives rejected

Tune the wording until it lands between the two. The gap being aimed at is between “worse than nothing” and “stops checking”, against a metric whose control arm has a range of 6 to 13 on three runs. There is no evidence a stable middle exists, and prompt-tuning against a noisy metric until a number improves is how a result gets manufactured.

Keep it off by default. A pillar nobody runs is not a pillar, and it still has to be maintained and reasoned about. The eval numbers would come from the path with it disabled.

Remember routes instead of failures. This was the plan’s original idea and the numbers foreclose it: eight of ten tasks are already at the floor, so there is no route knowledge worth carrying.

Consequences

• No agent state persists between sessions. Every run starts from what it can see, which is what the rest of the stack already assumes.
• The repeat-encounter harness goes with it. It exists in the commit below and is worth restoring if the question is reopened.
• The implementation, its eleven unit tests, and the harness are recoverable from commit f55e31e, “Build memory, and measure what it actually does”.
• What would reopen this: a task set where first-encounter exploration is genuinely expensive, such as an unfamiliar third-party app with deep navigation, where the control arm would show real discovery cost to recover. On this set there is five actions of it in total.
• The S2 attempt guard turned out to be doing more than its own job. It is the only reason the assertive arm was recorded as a failure rather than as a threefold improvement, and it is worth keeping in mind that eval guards catch things they were not written for.