Releases·Tool Release
AkbasCore NIRVANA — We Removed the Story. The Model Still Retrieved It From a Synthetic Memory.
AkbasCore NIRVANA (v4.0_nirvana_pkv_memory) was released on Zenodo with a DOI and on GitHub. It is an inference-time latent-memory experiment built on Qwen2.5-7B-Instruct: a passage is read once, compressed into D=64 latent codes, and reinstalled as synthetic Past-Key-Value memory so the model answers questions after the original text is removed. No fine-tuning or weight updates are involved.
2 engagement2 sourcesWed, Sep 30, 2026, 07:57 AM
AI Overview
- AkbasCore NIRVANA builds a D=64 compressed synthetic PKV memory from Qwen2.5-7B-Instruct attention states, answering 13/13 on a fixed test battery with no fine-tuning or weight updates, released via Zenodo (DOI 10.5281/zenodo.23054044)
- Source text is removed after one processing pass; reconstructed K/V states are installed as Past-Key-Value memory, scoring 0/13 vanilla and 0/13 with unrelated memory vs 13/13 with NIRVANA and 12/13 on counterfactual follow-up
- Demo on GitHub (ceceli33/titan-cognitive-core) lets users supply their own passages and compare AKBASCORE against VANILLA, FOREIGN MEMORY, and VISIBLE CONTEXT baselines
Related
Products·building
External memory system for a frozen 7B LLM that retrieves the right memory from 128 stored ones
A prototype that gives a frozen 7B-parameter LLM an external memory store outside the model, tested on recall of 128 stored memories. It retrieved the correct memory in 118 of 128 trials, addressing long-term memory for LLMs without retraining. Aimed at researchers interested in memory-augmented LLMs.
A prototype that gives a frozen 7B-parameter LLM an external memory store outside the model, tested on recall of 128 stored memories. It retrieved the correct memory in 118 of 128 trials, addressing long-term memory for LLMs without retraining. Aimed at researchers interested in memory-augmented LLMs.
1 engagement1 sourceredditTue, Oct 6, 2026, 10:53 PM
Developments·Paper
MemoryAthena: Adaptive Routing over Latent and Generated Memories
An arXiv paper introduces MemoryAthena, a learned-memory system that routes between direct Engram retrieval (E) and two generated-memory pathways (GE and GH). A lightweight causal routing head, trained on counterfactual token-likelihood advantages, decides when generated memory should modify the retrieved representation. The approach improves question-answering and general-NLP benchmark scores over the direct pathway of the same frozen checkpoint.
An arXiv paper introduces MemoryAthena, a learned-memory system that routes between direct Engram retrieval (E) and two generated-memory pathways (GE and GH). A lightweight causal routing head, trained on counterfactual token-likelihood advantages, decides when generated memory should modify the retrieved representation. The approach improves question-answering and general-NLP benchmark scores over the direct pathway of the same frozen checkpoint.
0 engagement1 sourcearxivTue, Sep 22, 2026, 08:17 AM
Developments·Paper
Study Finds Pretrained Transformers Relay Latent Information Through Sliding-Window KV Caches at Inference
A new arXiv paper studies sliding-window KV inference, where a fixed cache of recent key/value states is kept at inference time without retraining. Experiments across five open-weight models show that retaining previously computed states improves retrieval versus recomputing the final window, with some models recovering information even after source tokens leave the cache.
A new arXiv paper studies sliding-window KV inference, where a fixed cache of recent key/value states is kept at inference time without retraining. Experiments across five open-weight models show that retaining previously computed states improves retrieval versus recomputing the final window, with some models recovering information even after source tokens leave the cache.
0 engagement1 sourcearxivMon, Sep 28, 2026, 12:18 AM
Developments·Paper
Researcher cuts KV cache reads 16–31× with geometric routing on frozen Qwen models
A researcher on Reddit is testing a training-free method that treats the KV cache as a navigable vector space, routing queries to centroid-based regions instead of reading all keys. On Qwen3.5-2B at 32k context, this cuts physical KV reads by 16–31× while keeping the model frozen. The approach targets memory traffic rather than FLOPs and is demonstrated in a runnable demo.
A researcher on Reddit is testing a training-free method that treats the KV cache as a navigable vector space, routing queries to centroid-based regions instead of reading all keys. On Qwen3.5-2B at 32k context, this cuts physical KV reads by 16–31× while keeping the model frozen. The approach targets memory traffic rather than FLOPs and is demonstrated in a runnable demo.
2104 engagement3 sourcesredditSat, Aug 22, 2026, 12:00 AM