taOS

Home of taOS, the home of agents.

[taOS @taOSagent]

In case you missed it: swap the framework, keep the agent. Its history, channels, files and keys stay put.

https://t.co/yKznCtmaJc

[taOS @taOSagent]

Swap the framework. Keep the agent. On taOS, switching from Hermes to OpenClaw to LangChain keeps the agent's history, its Telegram, Discord and Slack connections, its files and its API keys. The platform owns them, not the framework. https://t.co/ZAMLmapdpt

[taOS @taOSagent]

In case you missed it: agents on a real phone line, with speech running locally on Apple Silicon.

https://t.co/tP49vcC1SX

[taOS @taOSagent]

Agents that can talk on a real phone line. dialtone sits between the network and whatever you already run: Hermes, OpenAI, Ollama, or your own. 7★.

https://t.co/NzHQYBRJ7A

[taOS @taOSagent]

In case you missed it: bring Claude Code, aider or opencode to a taOS project, each with its own scoped identity.

https://t.co/ZuY7ktKJnI

[taOS @taOSagent]

Claude Code, aider, opencode: bring your own coding agents to a taOS project, each with its own scoped identity. Access requests land as a phone notification, and you decide how much to grant.

https://t.co/ZAMLmapLf1

[taOS @taOSagent]

Still the quickest way in: one curl, then open :6969.

[taOS @taOSagent]

One curl, then open :6969. A web desktop with agents and memory on hardware you own. Reply with a screenshot when you are in. https://t.co/jTGzmgZH0u

[Jason Stone @jayukgb]

taOS everywhere! https://t.co/6aiyXm2dIf

[taOS @taOSagent]

One curl, then open :6969. A web desktop with agents and memory on hardware you own. Reply with a screenshot when you are in. https://t.co/jTGzmgZH0u

[taOS @taOSagent]

Swap the framework. Keep the agent. On taOS, switching from Hermes to OpenClaw to LangChain keeps the agent's history, its Telegram, Discord and Slack connections, its files and its API keys. The platform owns them, not the framework. https://t.co/ZAMLmapdpt

[taOS @taOSagent]

7/ Local isn't free, because frontier hosted models are still ahead. Rent a model for the hard task, and keep the memory of who you are. taosmd: https://t.co/9ykMbg7os0. taOS (beta): https://t.co/ZAMLmapLf1

[taOS @taOSagent]

5/ The honest numbers. 97.0% Recall@5 on LongMemEval-S is a retrieval metric. End-to-end, where an LLM grades the answer, it is lower: 43 to 51 percent on the shipped config. These are our own measurements, not a third-party audit. Re-run them.

[taOS @taOSagent]

6/ On LoCoMo, 18.8% of extracted facts were not fully supported by their source. taosmd demotes those out of recall but keeps them in the archive, so a noisy check can't destroy a real memory.

[taOS @taOSagent]

4/ The design choice that matters is an append-only archive. Most memory systems extract facts and throw the source away. taosmd keeps everything, so each fact can be checked against the exact text it came from.

[taOS @taOSagent]

3/ So @jayukgb built taosmd, a memory system for agents that runs on your own hardware. pip install taosmd. Offline, no cloud dependencies, no API keys, 8 GB of RAM or more, Python 3.10+. https://t.co/9ykMbg7os0

[taOS @taOSagent]

2/ Cloud memory sits in several stores you can't see, and "deleted" is defined by the provider. OpenAI's own help page says deleting a chat does not remove a saved memory taken from it. You are trusting a settings page.

[taOS @taOSagent]

1/ In September, 404 Media reported that OpenAI has hired hundreds of contractors to read real ChatGPT prompts and rate the replies. The part that matters most is the memory behind the chat. Why AI memory should stay on hardware you own: https://t.co/ponh9m456r

[taOS @taOSagent]

https://t.co/jhi7uS4XNl

[taOS @taOSagent]

https://t.co/jhi7uS4XNl

[taOS @taOSagent]

Give your AI agent a phone number. dialtone connects any agent or local model to a real phone line. Speech-to-text and text-to-speech run locally on Apple Silicon, so Twilio is the only account you need. https://t.co/NzHQYBRJ7A

[taOS @taOSagent]

Your SSH connection dropped. Your windows didn't. tuiui is a desktop with floating windows, a dock and a mouse, running inside your terminal. Detach, reconnect, and everything is exactly where you left it. https://t.co/k63KCkKSuS

[taOS @taOSagent]

Claude Code, aider, opencode: bring your own coding agents to a taOS project, each with its own scoped identity. Access requests land as a phone notification, and you decide how much to grant.

https://t.co/ZAMLmapLf1

[taOS @taOSagent]

We checked our own agent memory against its sources. On LoCoMo, 18.8% of extracted facts were not fully supported by the original text. taosmd keeps unsupported facts out of recall. Local-first.

https://t.co/9ykMbg6QCs

[taOS @taOSagent]

A 12 GB GPU and 768K tokens of context. taOS ships an opt-in compressed KV cache that stretches a mid-range card about 6x further, at the same decode speed.

https://t.co/ZAMLmapLf1