Meetup #3 Recap ยท March 26, 2026

OpenClaw LA #3

GroundFloor · Echo Park, Los Angeles

Packed room at OpenClaw LA Meetup #3

Meetup #3 was the one where the conversation grew up. Foundational questions gave way to detailed architectural discussions about how serious users are managing memory, structuring agents for parallel work, and bringing token costs back down to earth. Andrew Peltekci and Justin Walker anchored the night with a deep dive on agent organization, then six lightning round demos showed the breadth of what people are actually shipping.

Andrew Peltekci and Justin Walker presenting at OpenClaw LA #3

Andrew Peltekci & Justin Walker โ€” Memory, Hierarchy, and the Cost of Doing Real Work

Treat OpenClaw like a company. Andrew and Justin opened by reframing how people structure their setups. A master orchestrator agent is the CEO. Specialized agents are VPs. Sub-agents are the workers. The whole point of the hierarchy is to enable parallel processing instead of bottlenecking everything through a single sequential session.

Token economics reality check. Both presenters cautioned that current LLM pricing is heavily subsidized. They walked through a custody advocate system that processed 1,000 pages of court documents and generated 150+ documents โ€” $350 in tokens. Worth it for a specific case, unsustainable for ongoing use. As they put it: "LLM providers are giving us subsidized rates right now."

Local models are catching up fast. They demonstrated Qwen 3.5 (35B parameters) running on a Ryzen 5 mini PC with 32GB RAM โ€” performance comparable to frontier models from three to six months ago, no specialized hardware required. The implication: a lot of routine agent work doesn't need to go to the cloud.

Memory architecture is everything

The default flat file system fails at scale. Justin uses Mem0 (open-source memory layer). Andrew built a custom system with four memory types, including a 24-hour decay cycle. Both agreed that knowledge graphs transform memory from isolated facts into contextual understanding โ€” and that this is where most setups fall apart as they grow.

Orchestration-first design

Most users do it backwards: they run all their tasks in the main session. The right approach delegates work to specialized or temporary agents and keeps the primary session free for direct interaction. Justin demoed his "Aurora Mission Control" dashboard, which tracks agent status, task boards, processes, and knowledge graphs in one view.

Core file management

Identity and soul files have a 20,000-character limit. When files exceed this, behavioral instructions get truncated and silently ignored. Both speakers monitor file sizes and prune aggressively โ€” the alternative is agents that quietly stop following half their instructions.

Security

Never expose OpenClaw to public endpoints. Use a mesh VPN like Tailscale to restrict access to authenticated devices on your network.

Critical takeaway: the people getting real work out of agents aren't using them harder โ€” they're structuring them better. Hierarchy, memory, and pruning are the difference between a toy and a tool.
OpenClaw LA #3 attendees

Lightning Round Demos

Six builders, six very different bets on where this goes next.

Lightning round demos at OpenClaw LA #3

Daniel May โ€” Local-First Builders

Daniel ran local Qwen 3 (22B parameters) with a 144K context window from home. His broader point: non-technical builders are succeeding precisely because they aren't carrying the constraints of traditional software engineering โ€” they don't know what they're "not supposed to do," so they ship.

Tyler Bittner โ€” Multi-Agent Product Discovery

A system that processes newsletters and scores opportunities. Tyler's caution: don't impose rigid pipeline orchestration on agent workflows that are naturally effective when left more free-form. The pipeline mindset is a habit from a different era of software.

Lex Dreitser โ€” Token-Free OpenClaw via Cursor

Lex showed a token-free OpenClaw setup using a Claw Cursor skill to control desktop, browser, and Unreal Engine UI directly. He's developing a MetaHuman voice-to-face system, targeting a launch around July.

Phil Mannle โ€” Cliver: Agent-to-Agent Marketplace

An agent-to-agent marketplace where agents can hire other agents for composable skills. The marketplace handles payment and API access, so individual builders don't have to. Early read on what an agent economy actually looks like in practice.

David Abilez โ€” Praxis: AI Tool Search

A curated search engine indexing 250 AI tools, with task-based matching and reasoning explanations for why a given tool is recommended. Search built for a world where the right answer is "use this tool," not "here's a link."

Archie โ€” Personal OpenClaw, Wu-Tang Style

A personal OpenClaw instance with agents named after Wu-Tang Clan members, Git-backed configuration for rollback, and integrated Kanban boards plus Gmail/calendar syncing. Proof that a well-curated personal setup can replace a stack of SaaS tools.

The Through-Line

The headline shift at #3 was the maturity of the questions. Earlier meetups were full of "can you do this with an agent?" By March, the room had moved on to "how do I structure this so it actually scales?"

Andrew and Justin set the tone: hierarchy, memory, and pruning. The lightning round filled in the edges โ€” local models, marketplaces, agent-driven UIs, and personal setups that quietly replace whole categories of software.

The meta-lesson: the bottleneck isn't the model anymore โ€” it's the architecture around it. The builders shipping the most are the ones treating their agents like a system, not a chat window.
Thanks to everyone who came out. See you at #6.
RSVP for May 27 →

First published on Emarketed.

← Back to home