lachtan/nanobot-runtime
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dream: 2026-06-07 18:36, 4 change(s)

Browse Source 
[MEMORY] For Ollama Cloud `/detach` research tasks, explicitly specifying the model (e.g., `qwen35`, `gemini`) is more reliable than generic model-agnostic prompts

[MEMORY-REMOVE] "kimi-k2.6:cloud is blocked for nanobot agent deployment due to random Chinese output drift" — duplicates the more detailed Kimi K2.6 Chinese-output bug entry two lines above; the blocked-status conclusion is already implied by "critical risk for Czech use"

[MEMORY-REMOVE] "GLM-5.1:cloud has no known bugs and is proven across hundreds of nanobot agent turns" — "proven across hundreds of turns" duplicates the primary-model entry ("best sustained productivity over long sessions/hundreds of rounds"); merge unique fact ("no known bugs") into the primary-model line
This commit is contained in:
nanobot
2026-06-07 21:10:26 +02:00
parent 18c1407965
commit e5fd1a278d
2 changed files with 3 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
memory/.dream_cursor
View File

@@ -1 +1 @@
323 324

5
memory/MEMORY.md
View File

@@ -55,7 +55,7 @@ This file stores important information that should persist across sessions.
## Agent Model Selection ## Agent Model Selection
- Ollama subscription covers all discussed models; price/cost excluded from model comparison criteria - Ollama subscription covers all discussed models; price/cost excluded from model comparison criteria
- Prioritizes agentic performance, correct tool calling, and overall result quality - Prioritizes agentic performance, correct tool calling, and overall result quality
- Primary nanobot agent model: GLM-5.1:cloud (best sustained productivity over long sessions/hundreds of rounds, strong real-world agent benchmarks) - Primary nanobot agent model: GLM-5.1:cloud (best sustained productivity over long sessions/hundreds of rounds, strong real-world agent benchmarks, no known bugs)
- Alternative nanobot agent model for tool-heavy tasks: Qwen 3.5:cloud (397B variant) - Alternative nanobot agent model for tool-heavy tasks: Qwen 3.5:cloud (397B variant)
- Conservative fallback nanobot agent model: DeepSeek V3.2:cloud - Conservative fallback nanobot agent model: DeepSeek V3.2:cloud
- OpenRouter is pay-per-token alternative to Ollama subscription for model access - OpenRouter is pay-per-token alternative to Ollama subscription for model access
@@ -69,7 +69,6 @@ This file stores important information that should persist across sessions.
- For Czech use with Chinese-English models: always explicitly specify language in system prompt (critical for Kimi K2.6, recommended for GLM-5.1) - For Czech use with Chinese-English models: always explicitly specify language in system prompt (critical for Kimi K2.6, recommended for GLM-5.1)
- Kimi K2.6 supports `preserve_thinking` mode for multi-turn agent scenarios (retains reasoning content across turns) - Kimi K2.6 supports `preserve_thinking` mode for multi-turn agent scenarios (retains reasoning content across turns)
- minimax-m3:cloud is blocked for nanobot agent deployment due to empty tool result responses (ollama/ollama #16389) - minimax-m3:cloud is blocked for nanobot agent deployment due to empty tool result responses (ollama/ollama #16389)
- kimi-k2.6:cloud is blocked for nanobot agent deployment due to random Chinese output drift
- deepseek-v4-pro:cloud is blocked for interactive nanobot agent use due to 15.4 tok/s and 57s TTFT - deepseek-v4-pro:cloud is blocked for interactive nanobot agent use due to 15.4 tok/s and 57s TTFT
- deepseek-v4-flash:cloud is a viable nanobot agent alternative with 1M ctx, MIT license, and ~30-50 tok/s speed - deepseek-v4-flash:cloud is a viable nanobot agent alternative with 1M ctx, MIT license, and ~30-50 tok/s speed
- qwen3.5:397b-cloud is a viable nanobot agent alternative with multimodal support, 1M ctx, 201 languages including Czech, but has speed and accuracy tradeoffs - qwen3.5:397b-cloud is a viable nanobot agent alternative with multimodal support, 1M ctx, 201 languages including Czech, but has speed and accuracy tradeoffs
@@ -79,9 +78,9 @@ This file stores important information that should persist across sessions.
- nemotron-3-ultra:cloud released 2026-06-04 and is too new for real-world nanobot agent validation - nemotron-3-ultra:cloud released 2026-06-04 and is too new for real-world nanobot agent validation
- Czech language support is a hard requirement for nanobot agent models; Chinese output drift is a deployment blocker - Czech language support is a hard requirement for nanobot agent models; Chinese output drift is a deployment blocker
- GLM-5.1:cloud achieves ~198 tok/s on Ollama Cloud - GLM-5.1:cloud achieves ~198 tok/s on Ollama Cloud
- GLM-5.1:cloud has no known bugs and is proven across hundreds of nanobot agent turns
- Agent model comparison report saved to `results/2026-06-07_ollama-cloud-agent-model-comparison.md` - Agent model comparison report saved to `results/2026-06-07_ollama-cloud-agent-model-comparison.md`
- User prefers Qwen model for deep research tasks (not currently in presets) - User prefers Qwen model for deep research tasks (not currently in presets)
- For Ollama Cloud `/detach` research tasks, explicitly specifying the model (e.g., `qwen35`, `gemini`) is more reliable than generic model-agnostic prompts
- Available model presets: gemini-flash, gemini-flash-lite, glm, haiku, kimi, minimax, sonnet - Available model presets: gemini-flash, gemini-flash-lite, glm, haiku, kimi, minimax, sonnet
- minimax models have poor error recovery on tool calls — loop on blocked URLs instead of skipping and continuing - minimax models have poor error recovery on tool calls — loop on blocked URLs instead of skipping and continuing