Claude Memory Is Now Shared Across Chat and Cowork
Anthropic announced on Tuesday, August 25 that it has merged the memory systems behind Claude chat and Claude Cowork. What Claude learns in one now applies in the other. The practical effect is that the research phase and the execution phase stop being separate relationships — you can spend a week developing a plan in chat and then hand it to Cowork without re-explaining the headcount, the city, the speaker list or whatever else you already said.
The second change is quieter and probably more important: Claude now writes memories as topics during a conversation rather than summarising the whole thing when it ends. That closes the lag. Under the old model, anything you established mid-conversation was invisible to the other surface until the conversation was over. Now an ongoing chat can feed a Cowork session in something close to real time.
On the control side, Anthropic is exposing the memory list itself — users can read, edit or delete what Claude has retained on any topic. Sensitive categories are excluded by default: health data, race, ethnicity, religious beliefs, politics and gender identity are not stored unless you toggle on “include sensitive topics in memory,” and the app notifies you when a sensitive topic gets saved. Some categories are never stored at all — government-issued IDs, Social Security numbers, criminal history and immigration status among them. The feature is on by default for Free, Pro and Max across web, desktop and mobile; iOS and Android users need the latest app build.
Enterprises May Soon Keep the Mandatory 30-Day Logs on Their Own Cloud
Anthropic is planning a change to the data retention policy it introduced on June 9 for Claude Fable 5 and other models at that capability tier. The 30-day retention requirement is not going away — what changes is where the data sits. Enterprise customers are expected to be able to satisfy it on their own cloud infrastructure rather than Anthropic’s, via a new safety system due later this year.
This has been in development for months, reportedly with input from more than 100 customers including Salesforce. That customer count tells you how much friction the original policy created. A mandatory log of everything sent to a frontier model, held by the vendor, is a hard sell to any legal team that has already written a data residency clause into its own contracts.
Context worth holding: OpenAI previewed a zero-retention safety system last week, explicitly positioned against Anthropic’s requirement that logs exist. Anthropic’s answer is not zero retention — it is the same retention, under your key, in your account. Whether buyers read that as a meaningful concession or a repackaging is going to depend a lot on the implementation details, which have not been published.
Two Clean Days After Monday’s Multi-Model Incident
No incidents posted to the status page as of Wednesday morning. Third-party monitors checking through 11:03 UTC show Anthropic and Claude operational, with a small number of user-submitted reports in the trailing 24 hours — four for Claude, thirteen for Claude Code — which is background noise rather than signal at that volume.
That makes two consecutive clean days after the August 24 multi-model incident that ran 3 hours 26 minutes across Mythos 5, Fable 5, Opus 5 and Opus 4.8. August has been a rough month for uptime by this platform’s own recent standards, so a quiet stretch is worth logging rather than assuming.
Claude Code 2.1.246 Ships a Dedicated Agent Tool
After a week of narrow patch releases, 2.1.246 is a substantial one — roughly 61 CLI changes. The headline addition is a dedicated tool for launching agents on complex multi-step tasks, giving subagent workflows a first-class, isolated path instead of an improvised one. Each spawned agent gets its own context window, system prompt and tool access.
The second notable change is an auto mode tab in /permissions that lets you view and edit the auto-mode classifier rules. Auto mode has been the default permission mode for new sessions on Pro, Max and Team since August 14, and until now the classifier deciding what runs without a prompt has been effectively opaque. Being able to read those rules — and change them — is the difference between trusting a default and auditing one.
These two land together for a reason. Subagents have historically been the murkiest part of the permission story: they inherit modes, they can escalate, and they run with system prompts the operator did not write. A dedicated launch path plus inspectable classifier rules is the shape of an answer to that. Worth reading the changelog properly before you upgrade a fleet — 61 changes is not a patch you skim.
Session Budgets and the Playground Are the Two Console Changes That Stuck
Two August platform changes are worth revisiting now that the dust has settled. Managed Agents sessions can carry a budget — a hard cap on session spend, priced at public list rates, available since August 7. That is the control that makes an autonomous agent deployable in an org with a finance function, because “it will stop” is a materially different promise than “we monitor it.”
Separately, the Console Playground replaced Workbench, with full Messages API support, templates demonstrating code execution and web search, and the full SDK request plus API response shown for each run. The legacy Workbench and the experimental prompt tools APIs lost access on August 17 — if you had scripts hitting those endpoints, they are already broken, not about to break.
Thirty-Four Days Until Sonnet 4.5 Retires
Carrying yesterday’s item forward because the number only moves one direction. Claude Sonnet 4.5 retires September 29, 2026. After that date, API calls against claude-sonnet-4-5 fail — the same hard stop that hit Sonnet 4 and Opus 4 on June 15 and Opus 4.1 on August 5. Opus 4.5 stays.
If you did the grep yesterday, this is noise. If you did not, the job is unchanged: search every repo, config store, notebook, CI secret and vendor integration for the pinned identifier, not just the services you believe are live. Thirty-four days is comfortable. Six days is not.
Citigroup Joins the Top of the Syndicate as the Filing Window Narrows
Citigroup is set to join Goldman Sachs, Morgan Stanley and JPMorgan in the top ranks of advisers on Anthropic’s IPO. Citi currently holds the top spot in Bloomberg’s 2026 US IPO league table, largely on the strength of its role in SpaceX’s record $86.2 billion debut including the over-allotment — which is precisely the credential that matters if you are trying to run the deal that beats it.
Citi was also in the bank group behind Anthropic’s $2.5 billion revolving credit line last year, and that revolver is now set to rise above a roughly $10 billion target. Existing lender, top of the table, mega-deal experience — the selection reads as obvious in hindsight.
The company is reportedly weighing a public filing as soon as the end of this month, following the June 1 confidential S-1, ahead of a targeted October Nasdaq listing. Five days left in August. Nothing about ticker, price range or final valuation has been announced, and every revenue figure in circulation is still reporting rather than audited disclosure.
Ode with Anthropic Buys Casper Studios — 41 People Who Build Skills and Connectors
Ode with Anthropic, the $1.5 billion enterprise AI services venture backed by Anthropic, Blackstone, Hellman & Friedman, Goldman Sachs, General Atlantic and others that launched in July, has acquired Casper Studios. Casper is a 41-person, Select-tier partner founded in 2022 whose specific business is building the skills, connectors and context inside Anthropic products that make Claude useful in a given company’s systems. Terms were not disclosed.
The strategic logic is a split in the work. Ode does the large custom-engineered transformation programmes; Casper does the long tail of smaller, high-impact integrations that sit alongside them. Buying that capability rather than staffing it says something about how much of enterprise AI value now lives in the plumbing — the skill definitions, the connector configs, the institutional context — rather than the model.
Memory Is the Feature That Makes Switching Costs Real
Every AI vendor has spent two years claiming lock-in that does not exist. Prompts port. Model identifiers are a config change. The MCP spec means your connectors are increasingly portable by design. Until now, the honest answer to “what would it cost you to move off Claude?” was mostly retraining people and rewriting some glue.
Shared memory changes that answer, and it is worth being clear-eyed about how. Once Claude has accumulated months of topics about your projects, your preferences, your team’s shorthand and the specific way your org does things — and once that context flows automatically from the conversation where you think into the agent that acts — the cost of leaving is not a migration. It is amnesia. You do not port a relationship. You start one over.
Anthropic has handled the privacy side better than it had to. Sensitive categories off by default, an explicit toggle, a notification when something sensitive gets saved, hard exclusions for IDs and immigration status, and — the part that actually matters — a readable, editable, deletable list of what has been stored. That last item is the difference between a memory feature and a black box, and it is the one competitors most often skip. Credit where it is due.
But good privacy controls and accumulating switching costs are not in tension; they are the same product decision viewed from two angles. A memory you can inspect is a memory you are more willing to let grow. The controls make the moat easier to build, not smaller. That is not a criticism — it is a well-designed feature that happens to also be excellent business — it is just worth naming, because the pitch will be framed entirely as convenience.
The practical read for anyone deploying Claude at scale: treat the memory store as an asset with an owner. Know what is in it, know who can read it, know what your exit looks like. Ask the vendor now whether memory is exportable and in what format, before there is enough of it to matter. The teams that get burned here will not be the ones who thought too little about privacy — Anthropic has largely solved that for them. They will be the ones who never asked what happens to eighteen months of institutional context if the contract does not renew.