Aetna Hayes

AI - research / audit / trust, investing in AI, focused on turning research into products. pip install aetnamem

[Aetna Hayes @Aetna000]

Summary, in our controlled OpenClaw + DeepSeek benchmark, cache-aware AetnaMem beat native MEMORY.md on both:
Prompt tokens: 596,581 → 517,118, a 13.32% reduction
Provider-reported cost: 2.97% lower
Accuracy: 20/20 for both
Note: the improvement came mainly from sending a smaller, selective prompt not from increasing absolute cache hits. Results may vary by model, provider caching rules, memory size, and workload.

[Aetna Hayes @Aetna000]

We saw good saving on token but not cost in version 0.4.0 (https://t.co/F5SluKCRbP).

We improved it mad it memory that is cache-aware.

AetnaMem 0.4.1 is live. We separated stable memory from turn-specific recall and removed duplicate context. Result: 0.91% fewer tokens than our previous layout and 13.32% fewer than native MEMORY.md—with 20/20 accuracy. One Python/MCP contract now supports OpenClaw, Hermes and other agents.
https://t.co/GNoSQf6Q4d

[Aetna Hayes @Aetna000]

We did a benchmark comparing OpenClaw with and without Aetna Mem. We see about a modest 12-13% reduction of the token use, but it is not the whole story, there is cache that can make a difference.

Read the full report here: https://t.co/0S1GcHZAnD

Measured result:
40 real DeepSeek V4 Flash calls: 20 native-memory (memory.md) and 20 AetnaMem
Prompt tokens: 596,296 → 520,837
Reduction: 75,459 tokens, or 12.655%
Median per task: 29,808 → 26,028 tokens
Accuracy: 20/20 in both arms
Target memory retrieved: 20/20 AetnaMem trials
Audit chain: valid, covering 248 events

Conclusion : AetnaMem reduced total prompt context by 12.655% while preserving 20/20 accuracy, but cost was 0.674% higher because DeepSeek cached the large, unchanged MEMORY.md more effectively. AetnaMem’s query-specific recall blocks reduced reusable prefix caching, shifting some tokens from cheap cache reads to more expensive uncached input. In short, AetnaMem made the model context smaller and more relevant, but this warm-cache experiment shows that fewer tokens do not always mean a lower bill; selective memory and prompt caching must be optimized together.

[Aetna Hayes @Aetna000]

We are going to improve the design to make aetnamem cach-aware memory system

[Aetna Hayes @Aetna000]

We did a benchmark comparing OpenClaw with and without Aetna Mem. We see about a modest 12-13% reduction of the token use, but it is not the whole story, there is cache that can make a difference.

Read the full report here: https://t.co/0S1GcHZAnD

Measured result:
40 real DeepSeek V4 Flash calls: 20 native-memory (memory.md) and 20 AetnaMem
Prompt tokens: 596,296 → 520,837
Reduction: 75,459 tokens, or 12.655%
Median per task: 29,808 → 26,028 tokens
Accuracy: 20/20 in both arms
Target memory retrieved: 20/20 AetnaMem trials
Audit chain: valid, covering 248 events

Conclusion : AetnaMem reduced total prompt context by 12.655% while preserving 20/20 accuracy, but cost was 0.674% higher because DeepSeek cached the large, unchanged MEMORY.md more effectively. AetnaMem’s query-specific recall blocks reduced reusable prefix caching, shifting some tokens from cheap cache reads to more expensive uncached input. In short, AetnaMem made the model context smaller and more relevant, but this warm-cache experiment shows that fewer tokens do not always mean a lower bill; selective memory and prompt caching must be optimized together.

[Aetna Hayes @Aetna000]

Why We Need MCP When APIs Already Exist 🤔 https://t.co/f4RZm6GQt4 via @YouTube

[Aetna Hayes @Aetna000]

Like to know your opinion... Why this is misleading ....
It’s prompt caching, not memory ( we need to be careful about it !) Agree, we keep a large static prefix unchanged so the provider can reuse its computation, reducing repeated cost and latency by up to ~90%. The 108K to 11 tokens figure means only 11 tokens were new/uncached; the model still logically received the cached 108K context. BUT the cache expires (5 min or 1 hour max for Claude), exact-prefix changes can invalidate it (have they thought about it ?!); it remains provider specific too ... What you really need is a real memory 📝A real memory layer solves the larger problem that is persist knowledge across sessions, retrieve only the relevant evidence ... and work across Claude, OpenAI, Grok, and local models. A little bit biased here 📷 That is where AetnaMem fits. Of course there are others but my point is caching is not memory (1), it will expire and all the benefit is ephemeral in nature (2). What do you think?

[codila @0xCodila]

https://t.co/CAbiVW92hs

[Aetna Hayes @Aetna000]

This is a valuable optimization but it’s prompt caching, not memory ( we need to be careful about it !) Agree, we keep a large static prefix unchanged so the provider can reuse its computation, reducing repeated cost and latency by up to ~90%. The 108K to 11 tokens figure means only 11 tokens were new/uncached; the model still logically received the cached 108K context. BUT the cache expires (5 min or 1 hour max for Claude), exact-prefix changes can invalidate it (have you thought about it?); it remains provider specific too ... What you really need is a real memory 📝A real memory layer solves the larger problem that is persist knowledge across sessions, retrieve only the relevant evidence ... and work across Claude, OpenAI, Grok, and local models. A little bit biased here 📷 That is where AetnaMem (https://t.co/mGbJchMeRW) fits. Of course there are others but my point is caching is not memory (1), it will expire and all the benefit is temporal in nature (2).

[Aetna Hayes @Aetna000]

This is a valuable optimization but it’s prompt caching, not memory ( we need to be careful about it !)

Agree, we keep a large static prefix unchanged so the provider can reuse its computation, reducing repeated cost and latency by up to ~90%. The 108K to 11 tokens figure means only 11 tokens were new/uncached; the model still logically received the cached 108K context.
BUT the cache expires (5 min or 1 hour max for Claude), exact-prefix changes can invalidate it (have you thought about it?); it remains provider specific too ...
What you really need is a real memory 📝
A real memory layer solves the larger problem that is persist knowledge across sessions, retrieve only the relevant evidence ... and work across Claude, OpenAI, Grok, and local models.
A little bit biased here 😉 That is where AetnaMem (https://t.co/mGbJchMeRW) fits. Of course there are others but my point is caching is not memory (1), it will expire and all the benefit is temporal in nature (2).

[Aetna Hayes @Aetna000]

@m4npreet006 no question, Fable 5 is very good but many users got annoyed with the roller coaster they went through including myself.

[Aetna Hayes @Aetna000]

Fable 5 will remain one of the worst product releases in AI industry. It can be thought how not to ruin an amazing product

[Manpreet @m4npreet006]

Are you team fable 5 or chatgpt 5.6 right now ?

[Aetna Hayes @Aetna000]

Have you heard about #EtD (https://t.co/NOSQivvNds)?
Implemntation of EtD framework is now possible throught AetnaMem 0.4.0: a provider-neutral, auditable chain from evidence → EtD judgment → recommendation → approval → scoped authorization → guarded change.
PostgreSQL multi-user support, signed Ed25519/KMS receipts, COI/retention purge, offline verification, and a complete playground.
pip install aetnamem
https://t.co/rY8TLYLY1j
#OpenSource #ClinicalGovernance

[Aetna Hayes @Aetna000]

@elonmusk recently described a future of “universal high income” or material abundance driven by AI, robotics and solar energy.
I know @BillGates is not that popular these days. But just to remind that his thoughts here https://t.co/qBgd4HgWMh very insightful.

Indeed his pick about CO2 maybe outdated but I see merit in his discussion that can help us to move us forward.
Automation and low-cost energy could make a basic standard of living much cheaper, even when many scarce or customised goods continue to have prices.

[Aetna Hayes @Aetna000]

@elonmusk @tunguz Just for those interested to know more about "Innovating to Zero" https://t.co/rfQHwHXph9

[Aetna Hayes @Aetna000]

I recall Bill Gates once was saying that if energy becomes so cheap that we can produce it with almost 0 cost then we can get to eliminating the poverty and lack of resources… I think his idea at the time was investment on Nuclear Energy with high percentage of usage of the fuel rods instead of low percentage we have in older reactors.
I agree with Elon that everything can be free since we are moving towards lowering cost of energy production with sun and batteries, we are adding smart and automation in our daily lives that reduces the services costs.
The key is energy production and maintenance of the energy. If we reduce that the basics can be such low costs that will be free. It’s indeed gradual and dependent to the net zero and sustainability considerations since the complexity of the system we are building is going higher and higher the resilience.
The other factor is harmony, the world is connected and having actors that do not benefit from the system or are working against that can undermine it.

[Aetna Hayes @Aetna000]

https://t.co/O97KooFQ16