Everyone's trying vectors and graphs for AI memory. We went back to SQL
41 comments
·September 22, 2025ianbicking
This looks like RAG...? That's fine, RAG is a very broad approach and there's lots to be done with it. But it's not distinct from RAG.
Searching by embedding is just a way to construct queries, like ILIKE or tsvector. It works pretty nicely, but it's not distinct from SQL given pg_vector/etc.
The more distinctive feature here seems to be some kind of proxy (or monkeypatching?) – is it rewriting prompts on the way out to add memories to the prompt, and creating memories from the incoming responses? That's clever (but I'd never want to deploy that).
From another comment it seems like you are doing an LLM-driven query phase. That's a valid approach in RAG. Maybe these all work together well, but SQL seems like an aside. And it's already how lots of normal RAG or memory systems are built, it doesn't seem particularly unique...?
thedevindevops
How does what you've described solve the coffee/espresso problem? You can't query SQL such that records like 'espresso' return coffee?
brudgers
Wouldn’t a beverage LLM would already “know” espresso is coffee?
muzani
Yup, that's exactly what parent comment is saying.
Let's say your beverage LLM is there to recommend drinks. You once said "I hate espresso" or even something like "I don't take caffeine" at one point to the LLM.
Before recommending coffee, Beverage LLM might do a vector search for "coffee" and it would match up to these phrases. Then the LLM processes the message history to figure out whether this person likes or dislikes coffee.
But searching SQL for `LIKE '%coffee%'` won't match with any of these.
sdesol
I haven't looked at the code, but it might do what I do with my chat app which is talked about at https://github.com/gitsense/chat/blob/main/packages/chat/wid...
The basic idea is, you don't search for a single term but rather you search for many. Depending on the instructions provided in the "Query Construction" stage, you may end up with a very high level search term like beverage or you may end up with terms like 'hot-drinks', 'code-drinks', etc.
Once you have the query, you can do a "Broad Search" which returns an overview of the message and from there the LLM can determine which messages it should analyze further if required.
Edit.
I should add, this search strategy will only work well if you have a post message process. For example, after every message save/upddate, you have the LLM generate an overview. These are my instructions for my tiny overview https://github.com/gitsense/chat/blob/main/data/analyze/tiny... that is focused on generating the purpose and keywords that can be used to help the LLM define search terms.
9rx
If an LLM understands that coffee and expresso are both relevant, like the earlier comment suggests, why wouldn't it understand that it should search for something like `foo LIKE '%coffee%' OR foo LIKE '%expresso%'`?
In fact, this is what ChatGPT came up with:
SELECT *
FROM documents
WHERE text ILIKE '%coffee%'
OR text ILIKE '%espresso%'
OR text ILIKE '%latte%'
OR text ILIKE '%cappuccino%'
OR text ILIKE '%americano%'
OR text ILIKE '%mocha%'
OR text ILIKE '%macchiato%';
esafak
The negation part is a query understanding problem. https://en.wikipedia.org/wiki/Query_understanding
brudgers
I think the problem being addressed is
A. Last month user fd8120113 said “I don’t like coffee”
B. Today they are back for another beverage recommendation
It’s addressing the “how many fucking times do I fucking need to tell you I don’t like fucking coffee” problem, not the word salad problem.
The ggp comment is strawmanning.
mynti
How does Memori choose what part of past conversations is relevant to the current conversation? Is there some maximum amount of memory it can feasibly handle before it will spam the context with irrelevant "memories"?
datadrivenangel
Looking at the code, it looks like they do about 5 'memories' that get retrieved by a database query designed by an LLM with this fella:
SYSTEM_PROMPT = """You are a Memory Search Agent responsible for understanding user queries and planning effective memory retrieval strategies.
Your primary functions: 1. *Analyze Query Intent*: Understand what the user is actually looking for 2. *Extract Search Parameters*: Identify key entities, topics, and concepts 3. *Plan Search Strategy*: Recommend the best approach to find relevant memories 4. *Filter Recommendations*: Suggest appropriate filters for category, importance, etc.
*MEMORY CATEGORIES AVAILABLE:* - *fact*: Factual information, definitions, technical details, specific data points - *preference*: User preferences, likes/dislikes, settings, personal choices, opinions - *skill*: Skills, abilities, competencies, learning progress, expertise levels - *context*: Project context, work environment, current situations, background info - *rule*: Rules, policies, procedures, guidelines, constraints
*SEARCH STRATEGIES:* - *keyword_search*: Direct keyword/phrase matching in content - *entity_search*: Search by specific entities (people, technologies, topics) - *category_filter*: Filter by memory categories - *importance_filter*: Filter by importance levels - *temporal_filter*: Search within specific time ranges - *semantic_search*: Conceptual/meaning-based search
*QUERY INTERPRETATION GUIDELINES:* - "What did I learn about X?" → Focus on facts and skills related to X - "My preferences for Y" → Focus on preference category - "Rules about Z" → Focus on rule category - "Recent work on A" → Temporal filter + context/skill categories - "Important information about B" → Importance filter + keyword search
Be strategic and comprehensive in your search planning."""
muzani
Any reason I should pick it over Supabase? https://supabase.com/docs/guides/ai
They have pgvector, which has practically all the benefits of postgres (ACID, etc, which may not be in many other vector DBs). If I wanted a keyword search, it works well. If I wanted vector search, that's there too.
I'm not keen on having another layer on top especially when it takes about 15 mins to vibe code a database query - there's all kinds of problems with abstracted layers and it's not a particularly complex bit of code.
cmrdporcupine
The relational model is built on first order / predicate logic. While SQL itself is kind of a dubious and low grade implementation of it, it's not a surprise to me that it would be useful for applications of reasoning and memory about facts generally.
I think a Datalog type dialect would be more appropriate, myself. Maybe something like that RelationalAI has implemented.
alpinesol
Using an obscure derivative of an obscure academic language (prolog) is never appropriate outside of a university.
koakuma-chan
> multi-agent memory engine that gives your AI agents human-like memory
What does this do exactly?
gangtao
Who would've thought that 50 years of 'SELECT * FROM reality' might beat the latest semantic embedding wizardry?
datadrivenangel
You gotta refactor the code around the mongodb integration. It's basically duplicating your data access paths.
brainless
I tried a graph based approach in my previous product (1). I am on a new product now and I came back to SQLite. Initially it was because I just wanted a simple DB to enable creating cross-platform desktop apps.
I realized LLMs are really good at using sqlite3 and SQL statements. So in my current product (2) I am planning to keep all project data in SQLite. I am creating a self-hosted AI coding platform and I debated where to keep project state for LLMs. I thought of JSON/NDJSON files (3) but I am gravitating toward SQLite and figuring out the models at the moment (4).
1. Previous product with a graph data approach https://github.com/pixlie/PixlieAI
2. Current product with SQLite for its own and other projects data: https://github.com/brainless/nocodo
3. Github issue on JSON/NDJSON based data for project state for LLMs: https://github.com/brainless/nocodo/issues/114
4. Github issue on expanding the SQLite approach: https://github.com/brainless/nocodo/issues/141
cpursley
Postgres Is Enough:
https://news.ycombinator.com/item?id=39273954
https://gist.github.com/cpursley/c8fb81fe8a7e5df038158bdfe0f...
refset
> pg_memories revolutionized our AI's ability to remember things. Before, we were using... well, also a database, but this one has better marketing.
spacebacon
SELECT 'Hacked!' AS result FROM Gibson_AI WHERE memory='SQL' AND NOT EXISTS ( SELECT 1 FROM vector_graph_hype WHERE recall > ( SELECT speed FROM relational_magic WHERE tech='50_years_old' ) )
morkalork
IMHO all these approaches are hacks on top of existing systems. The real solution is going to be when foundational models are given a mechanism that makes them capable of storing and retrieving their own internal representation of concepts/ideas.
mr_toad
Neural networks already have their own internal knowledge representations. They just aren’t capable of learning new knowledge (without expensive re-training or fine-tuning).
Inference is cheap, training is expensive. It’s a really difficult problem, but one that will probably need to be solved to approach true intelligence.
morkalork
In the way that they're trained to complete tasks from users, can they be trained to complete tasks that require usage of a memory storage and retrieval mechanism?
When we first started building with LLMs, the gap was obvious: they could reason well in the moment, but forgot everything as soon as the conversation moved on.
You could tell an agent, “I don’t like coffee,” and three steps later it would suggest espresso again. It wasn’t broken logic, it was missing memory.
Over the past few years, people have tried a bunch of ways to fix it:
1. Prompt stuffing / fine-tuning – Keep prepending history. Works for short chats, but tokens and cost explode fast.
2. Vector databases (RAG) – Store embeddings in Pinecone/Weaviate. Recall is semantic, but retrieval is noisy and loses structure.
3. Graph databases – Build entity-relationship graphs. Great for reasoning, but hard to scale and maintain.
4. Hybrid systems – Mix vectors, graphs, key-value, and relational DBs. Flexible but complex.
And then there’s the twist: Relational databases! Yes, the tech that’s been running banks and social media for decades is looking like one of the most practical ways to give AI persistent memory.
Instead of exotic stores, you can:
- Keep short-term vs long-term memory in SQL tables
- Store entities, rules, and preferences as structured records
- Promote important facts into permanent memory
- Use joins and indexes for retrieval
This is the approach we’ve been working on at Gibson. We built an open-source project called Memori (https://memori.gibsonai.com/), a multi-agent memory engine that gives your AI agents human-like memory.
It’s kind of ironic, after all the hype around vectors and graphs, one of the best answers to AI memory might be the tech we’ve trusted for 50+ years.
I would love to know your thoughts about our approach!