Apple's agent memory paper says forget more: full transcripts scored worst

*Two vendors switched on chat memory the same day a study found that carrying everything forward hurts, and nobody has priced the risk of a shared memory store yet.*

Opening read

The most useful finding in today's evidence is that an AI agent handed its full history did worse than one with no memory at all, while an agent given a strict subset beat both. Memory is a filtering problem, and the vendors shipping it as a feature this week treat it as storage. Carry less between tasks, and let fewer people write to what you carry.

Agent memory: keep the spec, drop the reasoning

The consensus read is hard to argue with: agents fail on long work because they forget, so make them remember, and more is better. This week's products follow that logic. Anthropic folded Claude chat and Cowork into one interface right after upgrading the memory layer that carries user context across chats, with the merged product reaching Pro and Max subscribers first, and Sakana switched on cross-conversation memory in Sakana Chat the same day TechCrunch AI (https://techcrunch.com/2026/09/16/anthropic-merges-claude-chat-and-cowork-in-one-interface/) Sakana AI (https://sakana.ai/chat-fugumax/). A smart reader would conclude the race is to remember more.

A paper on Apple's machine learning research site, by Sanjana Pedada, Aditya Dhavala, and Neelraj Patil, argues the opposite. What survives between sessions in their design is the short list a person could check at a glance: what the task is, how the data is shaped, which tools are wired up, and what the output must look like. The reasoning the agent produced along the way gets thrown out. Across three enterprise scenarios that design completed 96% of tasks, against 79% with no memory, and the figure that should change your estimate is the third one: an agent handed its full history completed only 71% Apple Machine Learning Research (https://machinelearning.apple.com/research/shared-selective-persistent-memory). Remembering everything was worse than remembering nothing. The same system also has what the authors call a zero-token refresh, an update to recurring data that runs without calling the model at all, and it cut task time by 14 times. That refresh is also the only piece with a replication reported at all: it succeeded in all 12 trials across four public datasets, though nothing in the evidence says who ran that replication or whether it was independent.

The paper frames this as efficiency. My read is that it is a control result. What is worth keeping is what a person can inspect in a minute; what is worth discarding is what nobody would ever audit. That matters more once memory is shared, and the Apple design lets accumulated context be reused across users under role-based access (permissions tied to a person's job). A paper last month on memory injection showed that one poisoned exchange is enough to bend what an agent says later, and that the attacker never has to touch the store itself Hugging Face Daily Papers (https://tldr.takara.ai/p/2608.23471). That result covers one agent's later answers; whether a single write can compromise everyone who reads a shared store is the question I would test, and I would not assume it does until someone tries. I do not know how the 96% transfers outside the three scenarios the authors chose. What I would bet on is the shape of the result, which matches IBM's finding that memory's gains were larger on the strict measure of completing every variant of a scenario than on the loose one Hugging Face Blog (https://huggingface.co/blog/ibm-research/altk-evolve-hmm). On that evidence, memory's clearest payoff is finishing every variant of a task; whether it also lifts raw ability is a question those numbers do not settle.

For an engineer choosing what to persist, this is a design choice with a test attached; for a business leader deciding whether to switch on a vendor's memory for a whole team, it is a question of who can write to the shared store. Take one recurring agent workflow you already run and execute it three ways: no memory, the full prior transcript, and a hand-written page holding only the task spec, schema, tool settings, and output rules. Count completions. If the page beats the transcript in that workflow, you have a signal worth chasing: run the same comparison on the next workflow before you pay to carry history there. If the transcript wins, that workflow leans on prior reasoning, and that reasoning is the thing to protect.

Sovereign AI: two anchors, one still unbuilt

Sovereign AI (models built and hosted under your own country's rules) is now a product category with two anchors moving faster than their proof. Cohere and Aleph Alpha signed a definitive agreement to combine under the Cohere name, with dual headquarters planned for Berlin and Toronto; regulators still have to clear it, and the companies themselves put the closing somewhere later this year Cohere Blog (https://cohere.com/blog/cohere-and-aleph-alpha-sign-agreement). The same day Mistral's models went live behind Firefox's Smart Window beta in France and North America, with zero data retention a stated term for those Smart Window interactions and the UK and Germany promised for later this year Mistral AI (https://mistral.ai/news/mistral-x-mozilla/). One of these is a signed intention; the other is a switch somebody flipped. A business leader shopping for a sovereign vendor should sort the two accordingly: the combined Cohere does not exist until the closing notice, so any contract written against it needs a change-of-control clause, while the browser beta is live today and zero retention is already written into its terms for those interactions.

Guardrails: Washington waves them off, Brussels already enforces them

For a week this brief has tracked staff warnings about extinction risk from inside the big labs turning into a political fight. President Trump has now called those warnings a hoax and said the only guardrail AI needs is a strong and smart president Don't Worry About the Vase (https://thezvi.substack.com/p/trump-goes-full-hoax-on-ai-existential). That confirms my read that the warnings were a political event. What it does not settle is what Washington will actually do, and even the writer who flagged the hoax line reads Trump's position as something short of a demand for zero guardrails Don't Worry About the Vase (https://thezvi.substack.com/p/trump-goes-full-hoax-on-ai-existential). Brussels is a different story. The EU AI Office has held full enforcement powers over general-purpose models (the large models behind products like ChatGPT) since August 2 EU AI Act Newsletter (https://artificialintelligenceact.substack.com/p/how-the-ai-act-was-negotiated-and). A business leader drafting a compliance plan should build it around the European rules that already carry enforcement, and treat the American side as unsettled rather than absent.

Open models: the playbook is public, the call was a miss

This brief has recently covered cheap open-weight models, whose trained parameters anyone can download, as fallbacks to the frontier. Together AI now says some customers cut costs by up to 70% moving from closed to open models, and it recommends a canary (a small test slice) of 10% of traffic first Together AI Blog (https://www.together.ai/blog/migrating-from-closed-to-open-source-models). That is the playbook; the consumer version of the same bet shipped the same day. Sakana made Fugu Max, which routes prompts across a pool of open models and claims frontier parity, the default for every Sakana Chat user Sakana AI (https://sakana.ai/chat-fugumax/). On my own record: I called in August that Together, Fireworks, or OpenRouter would host DeepSeek-V4-Flash by September 15, and that call is graded a provisional miss. An engineer with a closed-model bill should run the 10% canary before believing the 70%.

Augur calls

• Call: Cohere publicly announces the completed closing of its business combination with Aleph Alpha on or before 2026-12-31.
• Criteria: Hit if Cohere or Aleph Alpha publishes an announcement, press release, or regulatory filing confirming the transaction has closed by the due date; miss if no closing confirmation is public by then or the deal is terminated.
• Due: 2026-12-31
• Odds: 60%
• Call: Mozilla or Mistral publicly confirms Firefox Smart Window with Mistral models is available to users in both the United Kingdom and Germany on or before 2026-12-31.
• Criteria: Hit if an official Mozilla or Mistral announcement, release note, or product page states Smart Window is live for UK and German users by the due date; miss if either country remains unsupported or the rollout is publicly delayed past it.
• Due: 2026-12-31
• Odds: 55%

What to watch

• The Cohere and Aleph Alpha deal still has to get past regulators before a closing later this year, and the closing notice is where the dual-headquarters plan gets tested. Cohere Blog (https://cohere.com/blog/cohere-and-aleph-alpha-sign-agreement)
• Mozilla's promised UK and Germany rollout of Smart Window will show whether the zero-retention terms hold outside France and North America. Mistral AI (https://mistral.ai/news/mistral-x-mozilla/)
• Anthropic's merged Claude is reaching Pro and Max users first; the thing to test is what memory carries between chat and Cowork. Simon Willison's Weblog (https://simonwillison.net/2026/Sep/16/one-claude/)
• Apple's completion rates still have no reported replication; only the zero-token refresh (the update that skips the model) does, and whether that run was independent is not something the evidence settles. Apple Machine Learning Research (https://machinelearning.apple.com/research/shared-selective-persistent-memory)
• The EU AI Office has held enforcement powers since August 2, and the first public action against a general-purpose model provider would set the real floor. EU AI Act Newsletter (https://artificialintelligenceact.substack.com/p/how-the-ai-act-was-negotiated-and)

Verification notes

The completion rates in the lead come from three enterprise scenarios chosen by the paper's authors, and the only replication reported covers the zero-token refresh mechanism, on four public datasets, with nothing in the evidence saying who ran it Apple Machine Learning Research (https://machinelearning.apple.com/research/shared-selective-persistent-memory). The memory injection result I have describes steering one agent's later answers from a single interaction; it says nothing about a store shared across many users. Together AI's 70% cost figure and Sakana's frontier-parity claim are both vendor self-reported, Firefox Smart Window is a beta, and the Cohere and Aleph Alpha combination has not closed and still requires regulatory approval Cohere Blog (https://cohere.com/blog/cohere-and-aleph-alpha-sign-agreement).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-17?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-17)