research-memory-gateway

name

research-memory-gateway

description

Guides agents to use the Research Memory Gateway MCP for evidence-first long-term memory across Codex, Cherry Studio, KiloCode, and related tools. Use when doing literature review, paper notes, synthesis route planning, experiment planning, mechanism hypotheses, material-system tracking, presentation outlines, research decisions, or reusable agent/MCP/deployment configuration lessons that may need long-term recall.

Research Memory Gateway

Use this skill when the user is doing reusable scientific research work, or when the session produces reusable agent/MCP/deployment configuration knowledge, and the research-memory-gateway MCP tools are available.

This skill is the portable cross-agent policy. New deployments should install or inject this skill for every agent/client that can access the MCP; do not rely on a client-local memory feature to write to this gateway automatically.

Core Rule

Long-term memory is for reusable research assets and reusable operating knowledge, not ordinary chat history.

Never save directly unless the user has explicitly confirmed. First call propose_save; call save_research_memory only after confirmation with user_confirmed=true.

When confirmation happens in chat, pass a confirmation payload so the gateway records that the save was user-confirmed and does not require a second WebUI approval:

{
  "source": "chat",
  "text": "<the user's confirmation text>",
  "confirmed_by": "user"
}

Be proactive about proposing saves. Loading the MCP tools or enabling a client's local memory does not make this gateway a memory backend; the agent must explicitly call the gateway tools when durable knowledge is produced.

When To Propose Saving

Call propose_save when the session produces one of these assets:

• A literature review or technical-route comparison.
• A reusable paper note.
• A synthesis route, precursor choice, reaction condition plan, or risk analysis.
• An experiment plan, control design, characterization workflow, or expected-result criterion.
• A mechanism hypothesis, falsification condition, or evidence evaluation.
• A material-system summary, performance table, or structure-property relation.
• A PPT, report, group-meeting, or thesis-defense outline.
• A research decision that changes future work.
• A reusable agent behavior rule, MCP integration lesson, deployment decision, client configuration pattern, or troubleshooting result that should apply to future agents.

Do not propose saving for small clarifications, one-off translations, casual discussion, or incomplete speculation unless the user asks to remember it.

For reusable agent/MCP/deployment configuration knowledge, usually use memory_type: workflow_plan / 工作流规划 with metadata.plan_type and metadata.plan_status, plus a project such as research-memory-gateway or the relevant client/project namespace.

Tool Workflow

1. If current work may duplicate earlier work, call check_overlap first.
2. Build a strong-evidence memory object.
3. Call propose_save with the reason and suggested memory.
4. Show the user a concise save proposal: type, title, claims, evidence count, source_refs, overlap candidates.
5. Wait for explicit confirmation.
6. If confirmed, call save_research_memory with user_confirmed=true, the proposal_id, and confirmation.
7. If the user asks to verify or trace a claim, call open_source_ref before relying on the memory.
8. If the user asks for memory hygiene, call audit_unverified.

Strong Evidence Memory Shape

Use exactly one of these memory_type values:

• literature_review / 文献综述
• paper_note / 论文笔记
• synthesis_route / 合成路线
• experiment_plan / 实验规划
• mechanism_hypothesis / 机制假设
• material_system / 材料体系
• presentation_outline / 汇报提纲
• research_decision / 研究决策
• workflow_plan / 工作流规划

For experiment_plan / 实验规划 and workflow_plan / 工作流规划, include metadata.plan_status / 规划状态:

• draft / 草案: not actionable by default.
• accepted / 已确认: user-confirmed and actionable.
• active / 执行中: currently in force and actionable.
• superseded / 已被取代: historical only.

For workflow_plan / 工作流规划, also include metadata.plan_type / 规划类型 when useful:

• agent_memory_policy / Agent 记忆策略
• mcp_setup / MCP 配置
• research_workflow / 科研工作流
• writing_workflow / 写作工作流
• deployment_workflow / 部署工作流
• project_governance / 项目治理

Every proposed memory should include:

• project: research project or topic namespace.
• topic: specific scientific topic.
• memory_type: one of the allowed types.
• title: short searchable title.
• summary: retrieval entry point only, not the source of truth.
• claims: checkable conclusions with confidence and verification status.
• evidence: exact quote, DOI, URL, paper title, file path, or source excerpt when available.
• source_refs: original session, file, DOI, or URL anchors.
• entities: materials, analytes, papers, methods, precursors, solvents, metrics, hypotheses.
• relations: lightweight graph edges between entities.
• tags: stable searchable keywords.
• next_actions: follow-up experiments, searches, or verification work.

Anti-Hallucination Rules

Treat summary as an index, not evidence.

Use evidence_backed only when a claim references valid evidence_ids.

Use unverified when evidence is missing.

Use inferred when the claim is reasoned from evidence but not directly stated.

Use conflicting, superseded, or retracted when older memories should not be used as-is.

Do not present an unverified or inferred claim as an established scientific conclusion.

Save Proposal Format To User

After propose_save, present this concise summary:

建议保存为长期科研记忆：
类型：<memory_type>
标题：<title>
项目：<project>
可核查结论：<n> 条
证据：<n> 条
来源锚点：<n> 个
潜在重复/冲突：<n> 条
状态：等待用户确认后才写入

Do not end with an open-ended offer. If confirmation is needed, ask one direct confirmation question.

Proposal lifecycle is separate from plan status. proposal_status / 提案状态 may be pending / 待审, approved / 已批准, rejected / 已驳回, needs_edit / 需修改, saved / 已保存, or expired / 已过期.

Retrieval Rules

When answering from memory:

• Prefer search_research_memory before relying on vague recollection.
• Report whether claims are evidence_backed, inferred, or unverified.
• Use open_source_ref when the user asks for provenance or when a high-stakes experimental decision depends on the claim.
• Surface conflicts instead of hiding them.