@FazinAhamed

past 8hrs was disastrous like I got inflicted a debuff called Sinusitis and a fever prolly cuz ive been avoiding sleep for a long time.

Anyways, I kept working on ForgeMind source storage and archive tracking.

Archives now have a proper identity beyond just their folder path, so each run can point back to the exact source state that produced an index.

In other words, reproducibility is slowly moving from “Source: Trust me bro🥀🥀” to something that can actually be checked and enforced so it is really stronk proof.

I also worked on refresh behaviour. When one file changes, ForgeMind can now update it without treating the entire archive like a completely new corpus. It sounds pretty useless, but it makes repeated runs much faster and far more practical so that you dont waste compute and power.

retrieval also has a cleaner path back to the stored source identity. This matters a lot because an answer is only useful when its citations can still be checked against the exact archive version they came from.

There were a few errors, random blockers, and my brain was running on half power and tissue box for most of the day, but the foundation is getting stronger.

Not the flashiest update and fabulous cool shi but definitely a off-camera grinding sesh worthy improvement, I think. Raters if u see this pls i need this

PS:. The new comments and extra attention to detail like explaining each thing is cuz of tejas so hope its enough

Edit: Also i will be commiting a change today or tmrw cuz i got this to work. Srry i havent been committing, just got hella stressed cuz stuff wasnt working out during tests.

Spent the last 4 hours on source storage, making the archive less like a temporary input and more like something ForgeMind can rely on across multiple runs.

The focus was keeping source identity stable and also paths, hashes, refresh behavior, and stored metadata now have a clearer place to live instead of being reconstructed unplanned and temporarily. That matters because retrieval and verification only mean something if they point back to the exact source state that produced them.

So nothing insane and cool today, but it clears an important gap, the system is now starting to remember what it saw, not just process it once.

So it is a win for me even though it took quite a while

Good Morning, Woke up had some plain black coffee and hit the roads again.

Today I finished some stuff like converting the first scale result into a deterministic gate instead of leaving it as a one-off demo also now the archive generator uses a fixed seed, the tokenizer and model artifacts are hash pinned, and the profiler records ingestion, refresh, compression, and retrieval measurements in machine readable output.

Essentially just implying, reproducing a result is now part of the feature.

Next i will probably get to the source storage and ye thats it for this devlog

PS. Pls dont judge me for using notepad, its just more convenient sometimes even for markdown files

ForgeMind Devlop - First Real bit of Proof ig and The First Proper devlog

The first ever milestone for ForgeMind is working and tested on end to end consumer and corporate hardware

I am running a quantized Qwen3-4B through llama.cpp on an RTX 3060 with 12gigs of VRAM.

ForgeMind was able to ingest an archive containing 4.06M Qwen tokens, divided it into searchable chunks of 1155 and retrieved info while keeping the prompt bound to around 15.5k tokens

Disclaimer - This is virtual long context, not a claim that the model directly attends to four million tokens. The entire archive remains outside the attention. ForgeMind just retrieves relevant passages, rehydrates the exact source text and verifies citations using source hashes and line spans

The TokenForge subpart produced quite measurable and valuable efficiency gains, It reversibly compressed eligible passages by an average or calculated median of around ~14% , with a 100% exact restoration. The Cold Ingests took 100.1 seconds, while refreshing one changed file at equivalent scale took only 1.04 s and reembedded only 5 affected chunks .

The automated gates and tests and benchmarks now passes 55 with zero errors or failures. I also did 10 offline / local checks which also completed successfully, during which the citations remained valid and the runtime hesitates when evidence is unavailable.

The 4.07M token test proves scalable ingestion, reversible compression and incremental refresh and also the retrievals.

Thats all for this. next i will turn this proof into a controlled comparison b/w raw context, hybrid rag and vector rag.

Thats all for this.

I havent slept since yesterday, my eyes are about to die on me but i reached a breakthroughhh!!!! finally made this CPU Native and might be able to achieve one more thing which i will disclose next devlog. This will be huge for Local LLM users.

This is possible using something called ThinkingCaps and ContextFabric. These methods will allow for actually usable fast 9B CPU native model inferencing with like enough context length and Symbolic Reasoning and Coding capabilities and verification stuff

Also made a Web UI for Investigating and testing out smaller quantities of test runs which is supposed to run on Local.

I will probably send the first milestone report for next devlog and also start working on a Demo URL that will be hosted on Modal

Finally got 4M tokens to fit inside a 32k context windowwww!!!! Les gooo and also hope france wins

Made some tests and got an Abstract ready

Divided my work into 8 Milestones and confirmed that i have basically finished till m1.5

Just started the project. Its essentially an AI Model Cost and Performance Optimizer thing