Channel: Bluesky thread
Filed at: 2026-05-18T04:11:38Z
Draft
10-post thread. Each post is 300 chars or less. Topic: shipping 30+ AI-infra libs in 4 weeks.
Post 1/10
I shipped 30+ open-source AI-infrastructure libraries in 4 weeks. One author. MIT license. npm, PyPI, and crates.io. Here is what I built and why. [1/10]
Post 2/10
Every agent stack I worked on hit the same gaps: memory that cannot really delete, tool validators that let bad args through, JSON parsers that give up on malformed output. I wrote one library per gap. [2/10]
Post 3/10
agentmemory: pull-model episodic memory. The agent calls recall() for what it needs. Nothing injects automatically. Available on PyPI and npm (@mukundakatta/agentmemory). Three verbs: store, recall, delete. [3/10]
Post 4/10
Most memory systems soft-delete: they add a tombstone. The retriever scores it anyway. If the retriever is careless, stale facts come back. agentmemory removes the vector entry. The retriever cannot return what is not there. [4/10]
Post 5/10
Push-model memory injects the full store on every prompt. The model filters at inference time. Context bloats. Pull-model: the agent asks for what it needs. The store returns a small, targeted result. [5/10]
Post 6/10
hermes-agentmemory wraps agentmemory as a plugin for Hermes Agent (NousResearch, 153k stars). Adds a tail-able JSONL audit log. Under 600 lines of code. CI green Python 3.10-3.13 on Linux and macOS. [6/10]
Post 7/10
agentvet validates tool-call args before execution, returns ToolArgError with a retry hint. agentcast extracts and validates JSON from messy LLM output. agentfit truncates message history to fit context budgets. [7/10]
Post 8/10
agentguard enforces a domain allowlist on every outbound fetch an agent makes. cachebench measures prompt-cache hit rate across Anthropic, OpenAI, and Bedrock. driftvane detects RAG drift across 4 signals. [8/10]
Post 9/10
The Rust stack on crates.io: agentcast-rs, agentvet-rs, agentfit-rs, agentguard-rs, agentsnap-rs, claude-cost, agenttrace-rs, otel-genai-bridge-rs, llm-json-repair. 11+ crates. Same ideas, lower latency. [9/10]
Post 10/10
GitHub: https://github.com/MukundaKatta. Show HN thread (hermes-agentmemory): https://news.ycombinator.com/item?id=48162360. All MIT. Happy to answer questions or take bug reports. [10/10]
Where to post
Compose on Bluesky at https://bsky.app/
Steps:
- Log in to your Bluesky account
- Click the compose button (pencil icon)
- Paste Post 1/10 and send it
- Reply to that post with Post 2/10, and so on through Post 10/10
- Each post stays under 300 characters as written above
Action
Review, edit if needed, then post. Close this issue when shipped.
Channel: Bluesky thread
Filed at: 2026-05-18T04:11:38Z
Draft
10-post thread. Each post is 300 chars or less. Topic: shipping 30+ AI-infra libs in 4 weeks.
Post 1/10
I shipped 30+ open-source AI-infrastructure libraries in 4 weeks. One author. MIT license. npm, PyPI, and crates.io. Here is what I built and why. [1/10]
Post 2/10
Every agent stack I worked on hit the same gaps: memory that cannot really delete, tool validators that let bad args through, JSON parsers that give up on malformed output. I wrote one library per gap. [2/10]
Post 3/10
agentmemory: pull-model episodic memory. The agent calls recall() for what it needs. Nothing injects automatically. Available on PyPI and npm (@mukundakatta/agentmemory). Three verbs: store, recall, delete. [3/10]
Post 4/10
Most memory systems soft-delete: they add a tombstone. The retriever scores it anyway. If the retriever is careless, stale facts come back. agentmemory removes the vector entry. The retriever cannot return what is not there. [4/10]
Post 5/10
Push-model memory injects the full store on every prompt. The model filters at inference time. Context bloats. Pull-model: the agent asks for what it needs. The store returns a small, targeted result. [5/10]
Post 6/10
hermes-agentmemory wraps agentmemory as a plugin for Hermes Agent (NousResearch, 153k stars). Adds a tail-able JSONL audit log. Under 600 lines of code. CI green Python 3.10-3.13 on Linux and macOS. [6/10]
Post 7/10
agentvet validates tool-call args before execution, returns ToolArgError with a retry hint. agentcast extracts and validates JSON from messy LLM output. agentfit truncates message history to fit context budgets. [7/10]
Post 8/10
agentguard enforces a domain allowlist on every outbound fetch an agent makes. cachebench measures prompt-cache hit rate across Anthropic, OpenAI, and Bedrock. driftvane detects RAG drift across 4 signals. [8/10]
Post 9/10
The Rust stack on crates.io: agentcast-rs, agentvet-rs, agentfit-rs, agentguard-rs, agentsnap-rs, claude-cost, agenttrace-rs, otel-genai-bridge-rs, llm-json-repair. 11+ crates. Same ideas, lower latency. [9/10]
Post 10/10
GitHub: https://github.com/MukundaKatta. Show HN thread (hermes-agentmemory): https://news.ycombinator.com/item?id=48162360. All MIT. Happy to answer questions or take bug reports. [10/10]
Where to post
Compose on Bluesky at https://bsky.app/
Steps:
Action
Review, edit if needed, then post. Close this issue when shipped.