ilya

git commit -m “watching football”

[ilya @iryaboy24]

@TheTailgateTent Check out https://t.co/rDOnqf3ZrG I’ve started putting together with Claude as a research agent

[ilya @iryaboy24]

@Schultz_Report Great list. Just missing CJ Carr

[ilya @iryaboy24]

@lukebogus This looks awesome.

[ilya @iryaboy24]

@sudoingX Qwen3.6-35B-A3B-UD-Q5_K_M.gguf \
--tensor-split 12,28 \
-ngl 99 \
-c 65536 \
-b 512 \
-np 1 \
--host 0.0.0.0 \
--port 8080 \
--flash-attn on \
--cache-type-k q8_0 \
--cache-type-v q8_0 \
--no-warmup

[ilya @iryaboy24]

@tobi I wrote something very similar locally. Persistent memory and Claude code hooks. https://t.co/9PI4CKryMY

[ilya @iryaboy24]

@LottoLabs 150 tokens per second with Qwen 3.5 35B. Can’t wait to try 3.6

[ilya @iryaboy24]

@garrytan Why would you use supabase instead of a local docker? Just plugging into another cloud SaaS for no reason

[ilya @iryaboy24]

@kimmonismus Gemm4 26B is absolutely amazing and run on a 5070ti 16GB. Can not wait to try 31B on a 3090 with a larger context and kv

[ilya @iryaboy24]

@NousResearch @Magnificode @openclaw Please check out the repo where I merge the 2 because I had the same question! https://t.co/FaBwr7IikR

Benchmark results have come out pretty good. And retrieval speed is as fast with no wasted tokens.

[ilya @iryaboy24]

@bcherny @trq212 @ClaudeCode @AnthropicAI @NousResearch @Teknium Benchmarks:
• 50.2% LongMemEval (beats Mem0, entirely local)
• 90.7% HaluMem false memory resistance
• 34ms avg retrieval

All self-hosted. No OpenAI. No cloud.

Star it, try it, break it 🐾
https://t.co/FaBwr7IikR

[ilya @iryaboy24]

I just open sourced my first project!

Shiba Memory, persistent, self-improving memory for AI agents.

34ms hybrid retrieval. ACT-R scoring. Native Claude Code + Hermes support. Fully local, no cloud deps.

🔗 https://t.co/FaBwr7IikR

Thread 🧵

[ilya @iryaboy24]

@bcherny @trq212 @ClaudeCode @AnthropicAI Hermes users, Shiba ships as a native memory provider plugin. 🔌

shiba_recall, shiba_remember, shiba_forget tools available to the LLM out of the box. Auto memory on every turn. Session summaries. Prefetch before each response.

@NousResearch @teknium

[ilya @iryaboy24]

What Shiba actually does:

• Remembers across ALL sessions & projects
• Hybrid semantic + full-text search
• Links related memories via a knowledge graph
• Instincts evolve into skills over time
• Ingests web, RSS, git, files
• HTTP gateway for any agent

[ilya @iryaboy24]

@sudoingX Adding a 3090 to the rig this week! I have a 5070ti and it runs gemma4 26B pretty well but not enough context bc of the space. Qwen14B is terrible. Plan to try gemma4 31B first on the 3090

[ilya @iryaboy24]

@Teknium Gemma 26B is the best one.
Hermes
↓
Gemma4 26B local
↓
Shiba Brain (living memory)
├── pgvector semantic search
├── Full-text hybrid search
├── Knowledge graph (memory_links)
├── ACT-R decay (memories that evolve)
└── HTTP gateway

[ilya @iryaboy24]

@Teknium Also make sure to create a cli wrapper so Hermes can talk to it and recall very quickly without using the full context window.

[ilya @iryaboy24]

@Teknium pgvector on docker has worked really well. Do a combination of what karpathy talks about with the wiki LLM obsedian md files and then underlying it with core/long term memory in the db

[ilya @iryaboy24]

@Teknium Can honcho run locally now as well?

[ilya @iryaboy24]

@potatokmish A solution is coming, https://t.co/vP1zQpdfTM