Short answer: It assigns each memory a comparable significance score so keep, consolidate, or fade decisions are deliberate.
Not every extracted fact deserves equal permanence; an allergy outweighs small talk from the same chat. A score is a graded signal, not only a binary keep-or-delete. LLMs can estimate importance directionally during extraction. Scores then drive cleanup, summarization priority, and decay. Engram can attach and act on importance per memory.
The previous chapter tallied what remembering everything actually costs a system. This chapter looks at the natural next question, once a system accepts that some memories deserve to be retained more strongly than others, how does it actually decide which is which. Importance scoring is the mechanism that turns that abstract goal into a concrete, comparable value attached to each individual memory.
Why Can’t a System Simply Treat Every Stored Memory as Equally Worth Keeping Forever?
Memories genuinely differ in how much they matter to a system’s ongoing usefulness, a customer’s stated allergy carries different long-term weight than an offhand comment about the weather made during the same conversation, yet both would be captured identically by a system that extracts facts without ever distinguishing their relative significance. Treating every memory as equally important isn’t actually neutral, it’s a decision, one that implicitly ranks a passing remark exactly as highly as a fact that genuinely deserves to survive far longer, simply by failing to make any distinction between them at all.
What Does It Actually Mean to Assign a Memory an Importance Score Rather Than Just Deciding to Keep or Discard It Outright?
An importance score expresses how significant a specific memory is on a continuous scale, rather than forcing every memory into a binary keep-or-discard decision at the moment it’s created. This distinction matters because significance isn’t usually obvious immediately, a memory that seems minor today might turn out to matter considerably later, and a continuous score lets a system weigh that memory appropriately against everything else competing for space or attention, rather than having irreversibly discarded it, or permanently enshrined it, based on an early, possibly wrong guess.
What Actually Determines How Important a Given Memory Should Be Considered in the First Place?
Several distinct signals typically feed into a genuine importance judgment, how emotionally or practically significant the underlying event seems, how directly it connects to a person’s stated goals or an agent’s ongoing task, and how much it’s been explicitly emphasized or reinforced since it was first captured. A memory describing a customer’s stated dealbreaker carries different weight than a memory describing a minor, one-off preference, not because one is inherently more “true” than the other, but because one is far more likely to matter the next time this specific customer’s data actually gets used for something consequential.
Can an Importance Score Actually Be Estimated Reliably by an LLM, Rather Than Requiring a Person to Manually Rate Every Memory?
Yes, and this is exactly what makes importance scoring practical to apply at scale rather than remaining a purely theoretical exercise. A language model can be prompted to rate how significant a specific piece of extracted information seems on a simple numeric scale, using the same kind of judgment it already applies when deciding which topic a fact belongs to during extraction. This doesn’t require perfect accuracy on every single memory to be genuinely useful, it only requires the scoring to be directionally right often enough that genuinely important memories consistently score higher than genuinely trivial ones across the system as a whole.
How Does an Importance Score Actually Get Used Once It’s Been Assigned, Rather Than Simply Sitting Alongside a Memory as an Unused Number?
An importance score becomes a genuine input to the retention and forgetting decisions covered throughout this Part, weighing which memories survive a cleanup pass, which ones get prioritized for consolidation into a higher-level summary, and which ones are allowed to fade or expire first when a system needs to reduce what it’s holding onto. A low-importance memory isn’t necessarily deleted immediately, but it becomes a natural first candidate when a system does eventually need to prune, while a high-importance memory earns the kind of durability that reflects how much it’s actually likely to matter down the line.
How Does Weaviate Engram Let a System Attach and Act On an Importance Score for Each Memory It Stores?
Weaviate Engram’s custom property scoping lets a system attach an importance value directly to each memory as structured metadata, making that value something later searches and cleanup processes can actually filter and sort by rather than a judgment that exists only in passing during extraction. Consider a personal-injury law firm’s case-management assistant, where some details from a client intake call matter enormously to the eventual case strategy while others are simply incidental color from the conversation:
from engram import EngramClient
client = EngramClient(api_key=os.environ["ENGRAM_API_KEY"])
client.memories.add(
"Client confirmed the other driver ran a red light, corroborated by an independent witness at the scene.",
properties={"case_id": "case-mvacc-5521", "importance": "high"},
)
client.memories.add(
"Client mentioned they were running late to pick up their child from school at the time of the accident.",
properties={"case_id": "case-mvacc-5521", "importance": "low"},
)
high_priority_facts = client.memories.search(
query="What are the key facts supporting this client's case?",
properties={"case_id": "case-mvacc-5521", "importance": "high"},
)
The corroborated red-light violation, genuinely central to proving liability, is scored and stored as high importance, while the incidental detail about the client’s schedule that day is scored low, letting the firm’s case-strategy searches surface the facts that actually matter to the case first, and letting any future cleanup or consolidation pass treat the incidental detail as a far more natural candidate to summarize away or let fade. This is exactly the value importance scoring delivers for a use case like personal-injury case management, where the facts that will actually decide a case deserve to stand out clearly from the much larger pool of conversational detail surrounding them.
Importance scoring gives a system a genuine, comparable basis for deciding which memories deserve durability and which ones are natural candidates to fade, replacing an implicit assumption that everything matters equally with a deliberate, examinable judgment. Importance is only one signal among several a system can draw on for this kind of decision, though, and two others deserve their own close look. Our next chapter, How do recency and access frequency guide retention?, takes up exactly those complementary signals.