Your AI has an audit log now (here's what it saw and did)
Buried in Microsoft's admin documentation for Copilot memory, under the heading "Is there audit logging for Copilot memory?", sits one of the most quietly damning sentences in enterprise software: "Memory and personalization actions don't generate audit log entries in Purview." Purview is Microsoft's compliance suite. Its whole job is recording what happened in your tenant. And the AI that builds a memory about each of your employees is exempt from it.
That sentence is not an outlier. I went through the current documentation for all four major assistants asking one question: when the AI does something on its own, remembering a detail, calling a tool, changing what it believes about you, is there a log? The short answer is no, not one you can see. This article is the long answer, plus what an actual AI audit log looks like, because I built one and shipped it.
Quick answer: No major consumer AI assistant shows you a log of what its memory or tools did. Copilot memory is explicitly exempt from Purview audit logging, per Microsoft's own docs. OpenAI's enterprise Compliance API logs authentication and admin events but has no memory-event category. Claude's Enterprise audit log covers about 31 event types, none of them memory. Gemini's admin logs record feature usage, not what was remembered. An AI audit log worth the name records every action with its input, source, and timestamp, and lets you revisit it later.
Why would an AI assistant need an audit log?
Because assistants stopped being chatbots. A modern assistant writes to your memory on its own, and with tool connections (MCP and the vendors' connector equivalents) it creates tasks, edits calendars, and queries your notes. Those are actions in your account, taken by software, sometimes without a prompt that you would recognize as authorizing them.
For any other system that acts on your data, the expectation is boring and settled: actions get logged. Your bank logs logins. Your document system logs access. The EU AI Act even wrote the principle into law, in Article 12: high-risk AI systems "shall technically allow for the automatic recording of events (logs) over the lifetime of the system." To be precise about scope, that requirement covers high-risk systems like hiring and credit scoring, not general-purpose assistants, and the EU's Digital Omnibus just pushed its deadline from August 2026 to December 2027. So no law currently forces ChatGPT to log its memory actions. But the direction of travel is written down: automated actions should leave records. Consumer AI memory is the gap the principle has not reached yet.
There is also a plainer reason. When your assistant asserts something wrong about you, or a task appears that you did not create, the first question is "what did the AI actually do, and why?" Without a log, that question has no answer. You are left rereading old chats and guessing.
What do the big assistants actually log?
I checked each vendor's current documentation. Here is the honest state of things as of mid-July 2026.
Microsoft Copilot
The bluntest case, because Microsoft documents the gap itself. Copilot memory is stored in a hidden folder in each user's Exchange mailbox. Purview retention policies do not apply to it, details inferred from chat history are "dynamic" and Copilot "might update or discard older details" on its own, and memory actions generate no Purview audit entries. The feature is still labeled a preview, so any of this could change. But today, the compliance suite that logs everything else in a Microsoft tenant does not log what the AI remembers about your staff. My compliance checklist article lists an audit trail as control number four for exactly this reason: it is the one most vendors fail.
ChatGPT
OpenAI sells a Compliance API to Enterprise customers. Its documented log categories are audit events, authentication events, and Codex events. Memory is not an event category. There is a separate compliance endpoint that lets admins query what memory currently contains, which is genuinely useful, but it is a snapshot of the present state, not a record of changes. And since the June 2026 Dreaming update, ChatGPT's memory rewrites itself in the background in a data layer separate from conversation logs, so even a full conversation export cannot reconstruct what the memory system did. One detail I find telling: OpenAI does keep a log of deleted memories, for up to 30 days, "for safety and debugging purposes." The log exists. You just cannot see it.
Claude
Anthropic's Enterprise audit log is real and reasonably broad: sign-ins, SSO changes, project and conversation lifecycle events, file uploads, around 31 event types. None of them are memory events. Your Claude memory updates roughly daily as a synthesized summary, and no tier shows you what changed or why. One honest caveat: Anthropic's Claude Managed Agents, a separate product for developers building enterprise agents, does ship a detailed memory audit log that records which agent and session a memory came from. So Anthropic clearly knows how to build this. It just is not part of the Claude that people mean when they say "Claude remembers me."
Gemini
Google Workspace admins get Gemini audit events, but they are usage events: which feature was used, which app, which file Gemini touched on the user's behalf. What Personal Context learned, updated, or quietly dropped appears in no log, admin or user. The consumer Activity page is a record of your prompts, which is chat history, not a record of memory changes.
The pattern across all four is the same: the memory page shows you the current state, in the model's own words, with no timestamps, no provenance, and no history. What it believed last month, and what changed it, is not on offer anywhere.
What about tool calls?
Here I want to be fair, because the situation is better than "no visibility."
When ChatGPT calls a connector or MCP tool mid-conversation, you can expand the call and see the actual request and response JSON. That is real transparency, and it deserves credit. The limitation is that it is ephemeral: it lives inside that one conversation, and there is no persistent, dated history of tool calls you can revisit across sessions. If you want to know what tools your assistant called last week, the answer is scrolling through every chat and hoping you find them.
Claude Desktop is rougher. Its MCP visibility is developer tooling: connection status in settings, and raw log files on your own disk if you know where to look. Fine for debugging a server you wrote. Not a way for a normal person to review what an assistant did in their account.
The distinction that matters is live-once versus revisitable. An audit log is not "you could have watched it happen." It is "you can check it later."
What does a real AI activity log look like?
Calmara is my product, so read this section as the pitch it is. It is also the concrete existence proof that this is buildable, on every tier, by a one-person company.
Every tool call an assistant makes against a Calmara account lands in an activity log the user can open in settings. Each entry shows whether the call succeeded, which tool was called, where the call came from (an MCP client like Claude Desktop, or Calmara's own assistant), how it authenticated, the timestamp, and the duration. Expanding an entry shows the exact input the assistant sent and a summary of what came back. Failures are logged too, including rate-limited calls, because an audit trail that only records successes is marketing.
The same idea repeats across the product's other automated surfaces, because tool calls are not the only thing an AI does on its own. Memory changes carry their own trail: every fact has a revision history with validity dates, source, and the quote it came from, which is what a time-travel query looks like. And the nightly scheduler writes a per-move log: which task moved, from which date to which, when. Three automated systems, three logs, all user-visible, none tier-gated. Auditability as a paid add-on would miss the point.
The honest tradeoff: a log this complete is only possible because Calmara is the substrate the assistant acts on, rather than the assistant itself. I cannot log what ChatGPT does elsewhere in your life. I can only promise that everything an AI does inside your workspace leaves a receipt.
Frequently asked questions
What should an AI audit log record?
At minimum: the action taken, the exact input, the source and authentication method, the timestamp, and the outcome, including failures. It should be persistent and revisitable, not just visible in the moment, and it should cover automated actions like memory writes, not just user-triggered ones.
Does any law require AI assistants to keep audit logs?
Not consumer assistants, today. The EU AI Act's Article 12 requires automatic event logging, but only for high-risk systems such as hiring or credit scoring, and the Digital Omnibus moved that deadline to December 2, 2027. The principle that automated actions need automatic records is in law; its reach just has not extended to general-purpose assistants yet.
Can enterprise admins see what Copilot's memory did?
No. Microsoft's documentation states that memory and personalization actions generate no Purview audit log entries and that Purview retention policies do not apply to Copilot memory. Admins can locate the hidden memory folder through eDiscovery, which shows contents, not a history of actions.
Doesn't ChatGPT already show tool calls?
During the conversation, yes, and the expanded JSON view is genuinely good. But it is per-conversation and ephemeral. There is no persistent history of tool calls across your account that you can consult afterward, which is what separates a display from an audit log.
Is there any assistant that logs memory changes?
For consumers, none of the big four. Anthropic's Claude Managed Agents, a developer product for building enterprise agents, does keep a memory audit log with per-session provenance, which shows the capability exists. It has not reached any consumer assistant's memory feature.
If you are evaluating an assistant for regulated work, the audit trail is one of six controls I would insist on before approval. The other five, and the exact questions to put to a vendor, are in the compliance controls checklist. And for what happens when you try to remove something an AI remembers, the companion piece to this one covers the difference between suppressing and deleting a memory.
Written by Dan Hagen