Everyone is keeping quiet about this for now, but it's something everyone will be using soon Claude + Jev = Second Brain Diogo Almeida, who developed the Jev system, shared a simple thought "The point of Jev isn't to make the model smarter in a conversational sense, but to... make it completely deterministic at decision points. If you have a choice of N actions, the system should yield an exact choice without hallucinations or wasted tokens" Action Plan: Jev decides what to keep, code enforces the established rules, and the language model is left strictly with writing the final note or answer At every input, Jev instantly determines its fate: save, set as a todo, skip, or flag for review, while evaluating privacy and its usefulness 30 days out Filter only allows data through if it crosses the threshold for value and confidence, keeping all private details strictly local on your device Instructs the model to generate a concise note of 5 to 12 lines focused on a single claim, prohibiting it from rewriting the core body to avoid lossy compaction Compares the new note against its neighbors, detecting duplicates, relationships, or contradictions, and automatically sets bidirectional links for highly relevant notes Pulls 10 to 20 candidates via BM25 or embeddings, then selects no more than 5 of the most precise notes to feed into the context window Runs a nightly cleanup by proposing duplicate merges, flagging outdated entries, and pinpointing conflicts, leaving the final call to you Integrates without building from scratch, running on your Obsidian vault, plugging into Claude Code, or acting as a ready-made memory framework Delivers decisions with 98.5% accuracy in just 0.30s, outperforming major frontier models and building memory 6.6x faster Your agent isn't forgetting because the model is weak; it forgets because nothing was deciding what was worth keeping in the first place Detailed instructions can be found in the article belowshow more

@Bober_smart defining clear inputs for Jev can make it more effective. I learned this when I hit a wall with my own AI project.

That is precisely how he solves the global problem

@Bober_smart haha, right? curious, what part stood out to you most?

I like the separation of responsibilities: deterministic rules decide what matters, while the model focuses on reasoning and communication. That could make agent memory much more reliable.

Yes, that is what makes this system special

bidirectional links saved my chaotic vault last month

Yeah, man

obsidian past - jev now

You should start using this as soon as possible

the best second brain might just be one that knows what to forget

By the way, yes, that's a great idea

Jev is a real treasure for the second brain

Yes, the second bridge looks completely different now

this could make AI memory much more useful

Yes, that is a more rational use

Retrieval is the easy half. What breaks my notes setup is stale entries piling up until the model trusts old decisions. Does Jev have any way to prune or expire what it wrote?

Ohh yes

This visualization slaps. Deterministic second brain is exactly what most agents are missing.

That separation could make agent workflows far more reliable

Yes, reliability is one of the main criteria

The way this terminal looks makes work feel good

Yes, this is a clear demonstration

but it still feels like a very complicated system to me

You only need to figure it out once, and everything will become clear to you

Soon everyone will start talking about it

Yes, you need to be one of the first

love the save vs skip approach

@flamurbahtijari

The division of labor decides

I've been personally taking some time to develop Jev systems for financial systems and I liked your idea, especially separating decision-making from generation. I’m curious where the determinism actually begins though. In finance, for example, an agent might read “management expects margins to decline 50 bps next quarter.” Whether that’s worth keeping depends heavily on the task and context, and later guidance could partially contradict it. Save/skip can be deterministic once the signals and thresholds are defined, but relevance and contradiction still require semantic judgment somewhere upstream. Would be interesting to see exactly what the 98.5% is measuring. Thanks!!

So real

second brain only worked for me when jev had a hard gate before write. otherwise it just filled notes with confident wrong summaries

we are similar😅

This design solves LLM hallucination perfectly, privacy local storage is also a plus.

this is a much smarter way to build memory

Second brain setups are getting good

it's a bold claim, but people often overhype new AI integrations without real use cases

Interesante separar la decisión de qué conservar del trabajo de redacción del modelo. Es una distinción parecida a la que estamos viendo en memoria para agentes:

deterministic decisions plus llm just writing is honestly the right split tbh