SuperLocalMemory

Agent memory with Laya, Jev + Google Gemma

Open Source

Developer Tools

Artificial Intelligence

GitHub

Visit WebsiteSee on Product HuntGithub ⧉Twitter ⧉Instagram ⧉

I'm one of the co-authors on the SuperLocalMemory 4.0 paper, so read this with that in mind.

What I'd like people here to see is that this didn't appear overnight. Varun has worked on this problem for about eight months, in the open, and each step came with a preprint:

1. February: SuperLocalMemory. Local-first memory for multi-agent AI, with a defence against memory poisoning through isolation and Bayesian trust scoring, and no cloud or LLM calls.

https://arxiv.org/abs/2603.02240

2. March: V3, the mathematical foundations. A retrieval metric from information geometry, memory lifecycle modelled as Langevin dynamics instead of hand-tuned decay, and a formal way to detect contradictions between memories.

https://arxiv.org/abs/2603.14588

3. April: V3.3, "The Living Brain". The full cognitive memory taxonomy, adaptive forgetting, and a distance metric that prefers full-fidelity embeddings over quantized ones.

https://arxiv.org/abs/2604.04514

4. August: V4, the governed memory system I worked on. Role-based access, verified erasure, a hash-chained audit trail, and a write path tested under eleven fault-injection scenarios, each repeated 200 times. It also opens with a negative result: ten mechanisms that were built and reachable but didn't work at their final connection, plus the checks we added to catch that class of bug.

https://arxiv.org/abs/2608.08253

Each paper added something the product actually ships, and the code is all open: https://github.com/qualixar/superlocalmemory

Happy to answer questions about the V4 governance and evaluation work.

I co-authored the SuperLocalMemory 4.0 paper with Varun, so take this with that disclosure in mind.

What I'd point technical readers to is how the paper handles its own results. It opens with a negative result: ten mechanisms were implemented, reachable on a live call path, and still ineffective at their final connection. "Implemented", "reachable" and "effective" turned out to be three separate questions, and the paper adds two mechanical checks to tell them apart.

It also tests the write path under failure: eleven fault-injection scenarios, each run 200 times, held 2,199 of 2,200 scoped properties. And it withdraws an overhead figure from the previous version because the two paths being compared weren't comparable. Measured properly, a governed write takes 11.0 ms, and most of that is the cost of durability, not governance.

Paper: https://arxiv.org/abs/2608.08253

Code: https://github.com/qualixar/superlocalmemory

Happy to answer questions on the evaluation or the reliability design.

I have seen the code, read the papers, this software is truly underrated. Performs better than most highly starred (github) memory repository. It's not based on somw glorified heuristic or strategy. Memory is stored in facts, facts are ranked appropriately, recency and relevan e is respected, old memories are forgotten. Wishing this launch its deserved success.

The distinction between a related memory and one that actually answers the question is the part that stood out to me. Returning 'no answer' seems much more useful than passing a vaguely relevant fact to the next agent.