What is the difference between semantic search and memory in AI systems?

Semantic search retrieves documents by vector proximity — it finds text that looks similar to your query but has no notion of state, contradiction, or time. Memory, by contrast, tracks relationships between pieces of knowledge, handles updates and contradictions, and evolves as the user interacts — giving your AI a persistent understanding rather than a one-shot lookup.

How does search result aggregation differ from re-ranking?

Re-ranking adds a reasoning step to select the most relevant documents from your top-K results, which improves precision but leaves recall capped at your search limit. Aggregation instead synthesizes multiple documents into each result slot, so a limit of 2 can still surface context drawn from 5 or more underlying memories — you get both precision and recall without raising your limit.

What happens to token usage and latency when you use aggregation instead of raising your top-K limit?

Aggregation keeps your search limit small, which directly reduces the tokens passed to your downstream LLM and the reasoning effort required — both lower cost and faster response times. Instead of feeding the LLM 20 raw documents hoping it finds the answer, you feed it 2 or 3 dense, pre-synthesized results that already contain the relevant context.

What is the precision-recall tradeoff in search, and why is it hard to solve?

Precision measures how relevant the returned results are; recall measures how much of the relevant information you actually retrieve. Raising your top-K limit improves recall but floods the LLM with noise, hurting precision. Lowering the limit sharpens precision but misses context, hurting recall. The two goals pull in opposite directions, and standard keyword or vector search offers no way to satisfy both at once within a fixed result budget.

Can I control how many aggregated results are returned, or is the output format fixed?

The aggregation feature respects whatever result limit you already set in your query — passing aggregate: true does not change the number of slots returned. Each slot simply becomes a synthesis of multiple underlying memories rather than a single raw document, so you keep full control over the size of the context passed to your LLM.

Should I use aggregation for simple single-turn queries or only for complex multi-part questions?

Aggregation delivers the most obvious value for complex, multi-part, or multi-session queries where a single document cannot fully answer the question — like 'What is this company and who runs it?' For straightforward single-fact lookups, standard search already performs well, and aggregation adds a synthesis step that may not be worth the overhead for that simpler case.

How does Supermemory's memory graph handle contradictions or updated information compared to a flat vector store?

A flat vector store indexes embeddings and returns whatever is most similar to your query — it has no mechanism to recognize that two documents conflict or that one piece of information supersedes another. Supermemory's memory graph tracks ontology-aware edges, handles knowledge updates and merges, and understands temporal ordering, so when a user's preference or fact changes, the graph reflects that update rather than returning stale context alongside current context.

How do I build a memory layer for my AI app without assembling five separate services?

Supermemory ships retrieval, memory, user profiles, connectors, and extractors as composable primitives in a single API — you can adopt one building block or wire together the full stack. The setup takes under 10 lines of code and integrates with LangChain, Vercel AI SDK, CrewAI, and other frameworks you're likely already using, so you're not ripping out your existing infrastructure.

What is SuperRAG and how does it handle retrieval differently from a standard vector search index?

SuperRAG is Supermemory's retrieval primitive — it runs hybrid vector and keyword search, applies context-aware reranking, and supports advanced metadata filtering and per-user scoping, all at sub-300ms latency. Unlike a plain vector index that returns the closest embeddings, SuperRAG integrates with the memory graph so retrieved documents carry relationship and temporal context, not just similarity scores.

What makes multi-session recall particularly difficult for standard RAG pipelines?

Standard RAG pipelines retrieve from a static document store with no awareness of what was asked across previous sessions, who the user is, or how their context has changed over time. When a query requires synthesizing signals from multiple past interactions — as most real-world personalized applications do — a fixed top-K retrieval window simply cannot hold enough slots to reconstruct the full picture, which is exactly the gap that aggregation and a persistent memory graph are built to close.

Solving the Precision-Recall Tradeoff in Search

When you're building memory for AI, search is your foundational layer. The way search generally works is straightforward: the user defines a query, and then sets a limit (top-K) on how many search results they want returned. Usually, this is set to 10 or 20.

Once the request hits the server, a response is returned sorted by semantic similarity. More often than not, this basic semantic match answers the query decently well.

However, there's a catch.

Semantic search fundamentally struggles with precision. Because it's just matching vector proximity, there's often noise in those top results. Just because a document uses similar words doesn't mean it actually has the context to answer the user's question.

The Problem with Re-ranking

To solve this precision issue, the industry introduced re-ranking. What re-ranking does is add a reasoning step—usually through a cross-encoder model or an LLM—to pass over the retrieved data and select what would actually be the best match to the query.

Re-ranking is fantastic for precision. It guarantees that the documents returned are highly relevant. But it suffers from a massive flaw: Recall.

You are still constrained by your search limit. If your limit is set to 3, but answering the user's question completely requires pulling context from 5 different documents, you are stuck. You simply do not have enough "slots" to fit the complete answer into the context window.

And if you're using an LLM anyways, might as well do the entire reasoning and only give the LLM what it needs.

The Supermemory Approach: Aggregation

We built Aggregation to solve exactly this issue.

What aggregation does is look at the same corpus of data, and it respects the exact same result limit. However, instead of one slot equalling one literal document, each result slot is a synthesis of multiple documents condensed into one.

By condensing multiple relevant memories to take up only one search result space, you no longer have to choose between precision and recall. You get both.

Seeing it in Action

Let's look at how this actually works within Supermemory. Imagine a memory space containing a bunch of unstructured data about our team and what we do.

Memory graph of hexagonal nodes linked to documents, with a version popup showing v1 memory entries and navigation

We want to ask a complex, multi-session query: "What is Supermemory and who is behind it?"

To illustrate the point, we'll set our search limit to just 2.

Without Aggregation: When we run this standard search, it returns two memories. These memories might match part of the question fairly well, but there just aren't enough slots to completely answer both parts of the question. The limit of 2 bottlenecks the recall, leaving the LLM with an incomplete picture.

───────────────────────────────────────────────────────────────────────
id: RbqTQjS85K88weNqZbAPzu
memory: Dhravya is the founder and CEO of Supermemory
similarity: 0.77
───────────────────────────────────────────────────────────────────────

───────────────────────────────────────────────────────────────────────
id: yHcHoGPpacSsMYxiKH4bdP
memory: Soham, Prasanna, Sreeram, and Tejas are part of the research team at Supermemory
similarity: 0.75
───────────────────────────────────────────────────────────────────────

With Aggregation: Now, we turn aggregate to true and run the exact same query with the same limit of 2.

It still returns exactly two results—but these results are aggregated from multiple underlying memories.

───────────────────────────────────────────────────────────────────────
id: aggregated_0_1775529845474
memory: Supermemory is an AI startup positioned as the context layer for the agentic era, founded by its CEO, Dhravya.
similarity: 1.00
───────────────────────────────────────────────────────────────────────

───────────────────────────────────────────────────────────────────────
id: aggregated_1_1775529845474
memory: The team behind Supermemory includes members across various functions: the research team consists of Soham, Prasanna, Sreeram, and Tejas; Mahesh serves as an engineer, and Shardul is part of the growth team.
similarity: 0.90
───────────────────────────────────────────────────────────────────────

Result 1 perfectly synthesizes multiple documents to answer the "What is Supermemory?" part of the query.
Result 2 aggregates entirely different memories to completely answer the "Who is behind it?" part.

Why this matters for your LLM apps

This feature is incredibly powerful for multi-session use cases. It allows developers to maintain a very small search limit—keeping downstream LLM token usage low, less reasoning effort, and latencies lightning fast—while still achieving the ridiculously high recall and precision required to answer complex questions.

You can simply turn on

aggregate: true

In the supermemory API to get started.

It's a huge step forward in building a memory engine that actually scales with human complexity.

If you want better memory for your LLMs and apps, you can use this today. Check out our docs, and get an API key on our dashboard.

Solving the Precision-Recall Tradeoff: Search Result Aggregation

The Problem with Re-ranking

The Supermemory Approach: Aggregation

Seeing it in Action

Why this matters for your LLM apps

Other posts.