Astrud

This was the breakthrough that made everything work. I discovered that you cannot expect an LLM to simultaneously have a concise, human-feeling conversational persona and rich tool-calling behavior. When I prompted a single agent to be concise like a human, it became concise with its tool calls too — refusing to do the heavy memory and execution work.

The solution: split them. An interaction agent handles conversation — its entire existence revolves around feeling human, matching the user's tone and verbosity. An execution agent works behind the curtain, processing delegated tasks with no constraints on thoroughness. When you send a long voice note, the interaction agent acknowledges instantly and delegates the heavy memory work to execution, so it doesn't block conversation.

I found this approach by discovering that Poke — an AI assistant whose conversational feel I admired — uses a multi-agent architecture. Someone convinced their Poke to email them its system prompts (plural). In hindsight, it's genius. The natural benefit is cost: the interaction agent can be a cheap, fast model. I'm using Grok 4.1 fast reasoning at 20 cents in / 50 cents out per million tokens. I can talk to it all day for basically free. Grok also gives me X search, which is useful for real-time information. The execution agent is Claude Opus 4.6. I was using Grok for execution too and wasn't getting the results I wanted. I started architecting a better system prompt, then thought — let me just plug in Opus and see if it just works. It just worked.

My first attempts at memory were a filing cabinet approach. I gave the agent tools — read_file, write_file — and access to files like Projects.md, Goals.md, Notes.md, Memory.md, User_Preferences.md, Personality.md. The idea was to give the agent freedom to organize its own internal state.

This didn't work. I was fighting two battles without knowing it. First, I was forcing the LLM to decide not only when to remember things, but how to categorize them. If you're an LLM and I tell you something, how do you decide whether it goes in Notes, Memory, or User Preferences? I was choosing these categories based on vibes — they were inherently unclear. Second, the concise persona problem was causing the agent to be concise with its tool calls, so it wasn't even trying to store memories properly.

I was discussing this with Opus 4.5, trying to architect a more human-like memory system. I remembered learning about human memory types in a college psychology course — short-term, working, long-term — and was trying to model encoding and surfacing. I asked Claude if there was active research on this, and it returned A-MEM — a paper out of Rutgers that presents a solution for creating a human-like memory bank for LLMs.

Here's how it works: when we want to remember something, we do an LLM call to format the memory as an atomic fact and generate tags for it. Then we generate a vector embedding and insert it into the brain. A smaller model analyzes whether the new memory node should be connected to existing nodes based on the embedding and the tags. At runtime, before the LLM responds to any query, we call surface_memories — we look at the user's message, get an embedding, and surface related memories from the brain. These get dynamically injected into the system prompt. Persistent, interconnected, atomic memories that surface automatically.

The first features I thought of were morning briefing — where I'm prompted with my active pursuits and held accountable to commit to working toward them that day — and evening briefing where I give a free-form recap of my day. This carries to weekly review where the LLM inferences my week and gives me a recap, then monthly, then yearly. Astrud doesn't just wait to be spoken to.

A full web dashboard built in React shows every LLM call, tool invocation, and token count as they happen via server-sent events. An Activity feed for raw event streaming. An Execution page for watching task processing in real-time with grouped tool calls nested under each LLM iteration. A Memory browser with live updates as new knowledge is stored.

The discovery that you can't have a concise persona and aggressive tool use in the same agent was the turning point. A single agent either feels robotic and verbose (optimized for execution) or feels human but does nothing useful (optimized for conversation). Splitting them means each agent's system prompt is entirely devoted to its job. The interaction agent's prompt literally says “IF YOU IGNORE EVERYTHING ELSE, ADHERE TO THIS” about matching user verbosity.

The filing cabinet approach — giving the agent categorized files to maintain — was rigid but unclear. The categories were vibes-based and the LLM couldn't consistently decide where information belonged. A-MEM eliminates the categorization problem entirely. Every memory is an atomic node. Organization emerges from vector similarity and tag connections, not predetermined buckets.

The interaction agent runs on Grok 4.1 fast reasoning — 20 cents in, 50 cents out per million tokens. I can talk to it all day for basically free. Grok also gives me X search for real-time information, which other LLMs don't have. The execution agent runs on Opus 4.6. I haven't done extensive price/performance testing, but Opus just works for complex multi-step tasks where cheaper models fall short.