Skip to main content

Future Research Summaries (Workspace-Relevant)

This page summarizes each link from /notes/future-research-links with practical takeaways for your workspace (OpenClaw workflow, KB, harness-engineering, multi-agent orchestration).

1) freeCodeCamp — Remembering preferences without breaking context​

Link: https://www.freecodecamp.org/news/how-to-build-ai-agents-that-remember-user-preferences-without-breaking-context/

Core idea: personalization breaks when you “just add more chat history.” Instead, split state into:

  • short-term context (belongs in prompt)
  • session state (scoped, structured)
  • long-term memory (durable preferences/facts)

Workspace-relevant takeaways

  • Matches what we’re already doing with MEMORY.md (long-term), memory/YYYY-MM-DD.md (daily logs), and project session logs (session state).
  • Reinforces the harness-engineering pattern: the agent is a planner, while the runtime/tools control execution (more debuggable, safer).
  • Suggests a clear policy: don’t let “memory” leak into every prompt. Retrieve only what’s needed (e.g., your “ask which project” preference).

Actionable next step for your stack

  • Add a small “Memory Policy” section to the Tutorial Coach Agent workflow:
    • what facts/preferences to store
    • when to ask permission
    • how to avoid cross-project contamination

2) Cloudflare — Markdown for Agents​

Link: https://blog.cloudflare.com/markdown-for-agents/

Core idea: serving Markdown directly (or via automatic HTML→Markdown conversion) is cheaper and cleaner for agents than parsing heavy HTML.

Workspace-relevant takeaways

  • Your KB is already Markdown-first (Docusaurus). This is basically confirmation that your “docs as source-of-truth” direction is future-proof.
  • If you ever put parts of your internal docs behind Cloudflare in the future, the “Accept: text/markdown” idea is a practical optimization for any agent that reads docs.

Actionable next step for your stack

  • In agent workflows where we fetch webpages, prefer markdown extraction (your existing web_fetch tool already does this). Keep “raw HTML” as last resort.

3) snarktank/antfarm — Multi-agent team workflows for OpenClaw​

Link: https://github.com/snarktank/antfarm

Core idea: a deterministic multi-agent pipeline (planner/developer/verifier/tester/reviewer) with explicit step ordering, retries, and verification gates.

Workspace-relevant takeaways

  • Very aligned with your “harness engineering” goal: explicit steps, separate verifier, reproducible runs.
  • This maps directly to your Tutorial Coach Agent model:
    • planner = decomposes tutorial steps
    • verifier = checks evidence + Q&A
    • reviewer = Arif approval
  • Antfarm’s emphasis on “fresh context every step” matches your project/session log approach (keep steps isolated, store state in files).

Actionable next step for your stack

  • Consider borrowing the concept of a workflow YAML for tutorials:
    • step ids: env → template → build → flash → monitor → menuconfig → Q&A
    • expects: evidence artifacts

4) Gist — “OpenClaw Implementation Prompts”​

Link: https://gist.github.com/mberman84/065631c62d6d8f30ecb14748c00fc6d9

Core idea: a set of long-form, copy/paste project prompts/specs for building systems (personal CRM, KB/RAG ingestion, idea pipeline, social research, etc.).

Workspace-relevant takeaways

  • These are essentially ready-made harness specs: clearly defined inputs, filters, dedupe rules, storage, and verification.
  • The “KB/RAG” section contains a strong ingestion checklist (fallback extractors, quality gates, dedupe, chunking, embeddings) that can influence your KB ingestion pipeline later.

Actionable next step for your stack

  • Create a folder to store these “spec prompts” in your repo:
    • docusaurus-kb/docs/notes/prompt-specs/ or
    • ~/.openclaw/workspace/prompts/specs/
  • Then adapt the “KB ingestion” spec into an SOP/project under ai-orchestration if you want a real ingestion tool.

5) sipeed/picoclaw — Ultra-lightweight assistant in Go​

Link: https://github.com/sipeed/picoclaw

Core idea: an assistant framework designed to run on very low-resource hardware (<10MB RAM) with a self-contained binary and its own workspace conventions.

Workspace-relevant takeaways

  • Interesting as inspiration for deploying “edge agents” on cheap hardware (SBCs) for monitoring/maintenance tasks.
  • Their documented workspace layout (sessions/memory/state/cron/skills + AGENTS/SOUL/TOOLS/USER) is conceptually similar to your OpenClaw workspace—good validation that your structure is sensible.

Actionable next step for your stack

  • If you later want an always-on agent on small hardware (for LAN monitoring), evaluate whether OpenClaw on a small VM is enough, or if a lighter agent runtime is worth it.

6) Hackster — ESP32 DataDisplay Terminal (CYD / esp32-2432s028)​

Link: https://www.hackster.io/news/your-desk-needs-an-esp32-datadisplay-terminal-0e1562b5c473

Core idea: Use a Cheap Yellow Display (esp32-2432s028, ESP32-WROVER + 2.8" 240×320 touch) as a desk “micro terminal” for glanceable data (weather/clock/etc.), with a 3D printed case and a browser-based flasher + OTA updates.

Workspace-relevant takeaways

  • This is a strong pattern for a desk-side status terminal for your home lab (VM health, ChirpStack gateway status, LoRaWAN metrics, build pipeline state).
  • The browser-based web flasher (Chrome/Edge) is a great distribution mechanism for workshop/team rollouts: no toolchain needed for initial install.
  • The device becomes self-service after flashing (touch UI for settings), which fits your “ops-friendly” direction.

Actionable next steps for your stack

  • Consider a “Desk Terminal v1” project that pulls from your existing portal endpoints:
    • portal.aurbotstem.com/api/vm-health/v1/status
    • (future) VM101 ChirpStack status endpoints
  • Evaluate whether to fork their firmware or build your own minimal UI using your existing ESP-IDF patterns.
  • freeCodeCamp memory article → tutorial-coach-agent + ai-orchestration
  • Cloudflare Markdown for Agents → ai-orchestration (web ingestion optimizations)
  • Antfarm → harness-engineering-team (workflow patterns)
  • Gist prompts → ai-orchestration (spec library)
  • PicoClaw → aurbotstem-infra (future low-footprint agent node idea)