awareliquid
A research team working on AI model architectures with episodic memory. It has proposed the O series and M series models.
Timeline (last 90 days)
Experiments published by the awareliquid team show that the O-series inference process holds memory constant at 0.381 MB when there are no session connections (at the 125M model scale).
In a 128k context measurement, the O-series' residual state memory is 1008 times smaller than conventional approaches.
The M-series uses a hybrid architecture that retains the attention mechanism, and its inference memory is not O(1).
The fast weight matrix stores key-value associations that have appeared across windows (episodic key-value memory), not a compressed distributed context.
The awareliquid experiment demonstrates that the fast weight matrix and its accompanying state can be serialized to disk across processes and restored bit-exact losslessly, with recall results after restoration identical bit-for-bit to the source process.
awareliquid retracted the '-28.5% adapter PPL' conclusion, attributing it to an error: the gain actually came from ordinary LoRA and had nothing to do with the liquid adapter.
awareliquid downgraded the notions of 'consciousness' and Phi values to sources of inspiration and no longer discusses them as metrics.
When evaluating 'it still remembers' demos, one should require all three elements: the storage form, the remaining performance after deleting the state, and cross-process restoration verification.
awareliquid achieves Recall@5 of 96.0% on LongMemEval, with zero LLM calls in the retrieval process.
The fast weight matrix does not bring understanding, reasoning or emotion, nor does it improve long-context language modeling performance, and the memory has no redundant backup.
All Facts (10)
Experiments published by the awareliquid team show that the O-series inference process holds memory constant at 0.381 MB when there are no session connections (at the 125M model scale).
50%UnverifiedIn a 128k context measurement, the O-series' residual state memory is 1008 times smaller than conventional approaches.
50%UnverifiedThe M-series uses a hybrid architecture that retains the attention mechanism, and its inference memory is not O(1).
50%UnverifiedThe fast weight matrix stores key-value associations that have appeared across windows (episodic key-value memory), not a compressed distributed context.
50%UnverifiedThe awareliquid experiment demonstrates that the fast weight matrix and its accompanying state can be serialized to disk across processes and restored bit-exact losslessly, with recall results after restoration identical bit-for-bit to the source process.
50%Unverifiedawareliquid achieves Recall@5 of 96.0% on LongMemEval, with zero LLM calls in the retrieval process.
50%UnverifiedThe fast weight matrix does not bring understanding, reasoning or emotion, nor does it improve long-context language modeling performance, and the memory has no redundant backup.
50%UnverifiedWhen evaluating 'it still remembers' demos, one should require all three elements: the storage form, the remaining performance after deleting the state, and cross-process restoration verification.
50%Unverifiedawareliquid retracted the '-28.5% adapter PPL' conclusion, attributing it to an error: the gain actually came from ordinary LoRA and had nothing to do with the liquid adapter.
50%Unverifiedawareliquid downgraded the notions of 'consciousness' and Phi values to sources of inspiration and no longer discusses them as metrics.
50%