How does consent affect what memory should remember?

Every chapter so far in this Part has assumed a memory was already appropriate to capture, and focused entirely on isolating it correctly once it existed. This chapter steps back a stage earlier, to a question that comes before isolation entirely: whether a given piece of information should have been remembered at all, and what role consent plays in answering that.

Why Isn’t Capturing Everything a User Says the Safest, Most Complete Approach to Building Memory?

It’s tempting to treat more captured information as strictly better, reasoning that a system can always choose not to retrieve something later even if it stored everything now. But this reasoning skips over a genuine ethical and legal distinction between storing information a person reasonably expected to be remembered and storing information they never actually consented to having preserved at all. A passing remark made in the flow of a conversation isn’t automatically the same thing as a durable fact someone knowingly agreed to have kept, and treating every utterance as fair game for permanent memory ignores that difference entirely.

What Does It Actually Mean for a Piece of Information to Be “Rememberable” in the First Place?

Rememberable information is information a person would reasonably expect, or explicitly agree, to have retained beyond the immediate moment it was shared. A user telling a support agent their account number to resolve a specific issue reasonably expects that number to be used for the current conversation, not necessarily to be extracted and stored as a durable memory retrievable in every future interaction. Treating “was this said in my presence” as sufficient grounds for permanent storage collapses a distinction that actually matters, between information shared for an immediate purpose and information a person meant, or was asked, to have remembered long-term.

How Does Consent Actually Change What a Memory System Should Be Configured to Extract in the First Place?

A well-designed memory system doesn’t extract everything indiscriminately and sort out consent later, it configures what it looks for in the first place based on what users have actually agreed to have remembered. This is precisely why defining a system’s topics deliberately, rather than configuring an overly broad, catch-all extraction target, matters for reasons beyond mere organization, an narrowly scoped topic naturally limits what gets captured to categories a user was actually informed about, rather than opportunistically extracting anything that happens to be extractable from raw conversation data.

Why Does the Right to Have Memory Deleted Matter as Much as the Right to Control What Gets Remembered in the First Place?

Consent isn’t a single decision made once and then permanent, someone who initially agreed to have certain information remembered can later withdraw that agreement, and a memory system needs a genuine, reliable way to honor that change of mind. This is exactly what a right to deletion protects, the ability for a person to have specific memories, or all of their memories, permanently and verifiably removed on request, not merely marked as inactive or quietly excluded from future searches while still technically sitting in storage somewhere.

What Actually Makes a Deletion Request Trustworthy Rather Than Merely Symbolic?

A trustworthy deletion actually removes the underlying data permanently, in a way that can be verified afterward, rather than simply hiding it from ordinary search results while leaving it recoverable somewhere in the system. A person who requests their data be deleted and later confirms that a search for their own information returns genuinely nothing has real evidence the deletion actually worked, which is a meaningfully different guarantee than trusting that a flag was set correctly somewhere in application logic. This distinction matters especially for a memory system precisely because memory, by its nature, is designed to persist, so a deletion mechanism has to work directly against that natural tendency rather than merely suppressing a memory’s visibility.

How Does Weaviate Engram Support Both Deliberately Scoped Consent and a Genuine, Verifiable Right to Deletion?

Weaviate Engram’s topic configuration lets a project define precisely what kinds of information get extracted in the first place, matching what users have actually been informed will be remembered, and its deletion API permanently removes specific memories or a user’s entire history on request. Consider a mental wellness journaling app where users share deeply personal reflections, some of which they explicitly agree to have the app remember as long-term context, and others they type simply to think out loud without ever intending them to be retained:

from engram import EngramClient

client = EngramClient(api_key=os.environ["ENGRAM_API_KEY"])

user_memories = client.memories.search(
    query="wellness goals",
    user_id="user-wellnessapp-5510",
)

for memory in user_memories:
    client.memories.delete(
        memory.id,
        user_id="user-wellnessapp-5510",
    )

remaining = client.memories.search(
    query="wellness goals",
    user_id="user-wellnessapp-5510",
)

The app’s topics are configured narrowly around goals the user explicitly agreed to have tracked over time, rather than indiscriminately extracting every personal reflection the user ever typed, and when this user later exercises their right to have that tracked history removed, the deletion call permanently erases each specific memory, with the follow-up search confirming genuinely nothing remains rather than merely trusting the deletion happened. This is exactly the value deliberate consent-aligned scoping and verifiable deletion deliver for a use case like mental wellness journaling, where the personal, sensitive nature of what users share makes both what gets remembered and what can be permanently forgotten matter just as much as how well any of it gets retrieved.

Consent reframes memory as something a person actively agrees to rather than something a system simply collects by default, and a genuine right to deletion makes that agreement meaningful rather than symbolic. Once a system respects what should and shouldn’t be remembered, a related question follows naturally: being able to trace, for any given memory, exactly where it came from and why it was captured in the first place. Our next chapter, What is auditability for memory origins?, takes up exactly that traceability.