Research Agent

The coordinator initializes the draft, routes work, and decides whether to iterate or move forward.

Literature review loop (stages 1-3)

/spawn research-agent.literature
/spawn research-agent.critique

Iterate until critique score ≥ 7/10.

Generate research brief

/spawn research-agent.brief

Review brief. If approved → proceed to baseline.

Implement baseline

/spawn research-agent.baseline

Shared Documents

Primary shared files:

• shared/research-draft.md — main source of truth
• shared/scorecard.md — candidate ranking and triage
• shared/experiment-plan.md — experiment specification
• shared/review-report.md — analyst + reviewer judgments
• shared/final-writeup.typ — final human-readable research writeup in Typst

Paper-Writing Standard

When the workflow is producing a paper, proposal, or paper-style draft, default to professional academic writing, not outline-style notes.

Required standard:

• Write dense, polished, publication-style prose.
• Introductions must motivate the problem clearly, explain why it matters, identify the gap, and frame contributions explicitly.
• Problem statements must define the setting, constraints, action space, objective, and why the problem is nontrivial.
• Related work should usually be written as flowing academic prose unless the user explicitly wants categorized subsections.
• Proposed methods must be detailed enough that a technically strong reader can understand what is being proposed and how it would work.
• Add mathematical formulation where needed, especially for objectives, constraints, and theory/error propagation.
• Citations and references are mandatory in paper drafts.
• Baselines and experimental framing must be concrete: specify candidate models, datasets/tasks, device classes, network conditions, metrics, staged experiments, ablations, and kill tests.
• Clarify ambiguous paper terms in the text (for example, what a deadline means operationally).
• Include figures/diagrams where useful; prefer self-contained generated figures if external assets are unnecessary.
• Keep claims scoped and defensible; do not fabricate results.

Default workflow for paper production:

1. Draft in markdown or Typst.
2. Run reviewer critique.
3. Revise to address reviewer comments.
4. Repeat reviewer/revision loop until major concerns are addressed.
5. Produce final Typst draft and compile a PDF when feasible in the environment.

If the environment cannot support a preferred template, fall back to a plain Typst layout that still compiles and preserves citations, figures, math, and professional formatting.

Agent Definitions

Quality Gates

Each stage has a score threshold. If critique scores below 7/10, iterate through idea → literature → critique until the direction strengthens.

Requirements

• Web search capability (brave search or similar)
• File write access to shared research draft
• Spawn permission for subagents

Memory Best Practices

For this pipeline to work reliably:

1. Keep workspace files small — total < 8KB. Agent prompts inject every message.
2. Write before compaction — use memory/YYYY-MM-DD.md for daily logs; MEMORY.md for durable facts. Rules in chat don't survive compaction.
3. Enable memory flush — set compaction.memoryFlush.enabled = true so the agent saves context before it gets trimmed.
4. Use memory_search — semantic recall over file-based memory, not mental notes.
5. Subagents for heavy work — keeps main session context lean.

Config Recommendations

{
  "agents": {
    "defaults": {
      "contextPruning": { "mode": "cache-ttl", "ttl": "5m" },
      "compaction": {
        "reserveTokensFloor": 20000,
        "memoryFlush": { "enabled": true, "softThresholdTokens": 4000 }
      }
    }
  },
  "memory": {
    "backend": "qmd",
    "qmd": {
      "includeDefaultMemory": true,
      "update": { "interval": "5m", "debounceMs": 15000 },
      "limits": { "maxResults": 6, "maxSnippetChars": 700 }
    }
  }
}

QMD Setup (Optional — for hybrid BM25 + vector search)

QMD is a local sidecar that replaces the default SQLite memory backend with better retrieval.

Prerequisites:

• QMD CLI: bun install -g https://github.com/tobi/qmd or npm install -g @tobilu/qmd
• Or use the npm global install already available

Initial setup:

# Create collection for your workspace
export XDG_CONFIG_HOME="$HOME/.openclaw/agents/main/qmd/xdg-config"
export XDG_CACHE_HOME="$HOME/.openclaw/agents/main/qmd/xdg-cache"
mkdir -p "$XDG_CONFIG_HOME" "$XDG_CACHE_HOME"

# Add workspace to QMD
qmd collection add ~/.openclaw/workspace --name "workspace"

# Generate embeddings (one-time, downloads ~300MB model)
qmd embed --collection workspace

# Verify
qmd status
qmd query "your search term" --json

What QMD provides:

• BM25 keyword search + vector similarity combined
• MMR (Maximal Marginal Relevance) reranking for diverse results
• Local GGUF embeddings via node-llama-cpp (no external API)
• Falls back to built-in SQLite if QMD is unavailable

When QMD is unavailable: OpenClaw auto-falls back to builtin SQLite manager — memory tools keep working.

Full Memory Layer Stack