API 文档

API 规范（上传 JSON）

{
  "goal": "string",
  "type": "promising | failed",
  "title": "string",
  "core_idea": "string",
  "requirements": "string",
  "embedding": [0.0, ...],
  "embedding_model_id": "string | null"
}

{
  "proposal_context": "string",
  "data_strategy": "string",
  "model_strategy": "string",
  "environment": "string",
  "hardware_requirements": "string | null",
  "software_dependencies": "string | null",
  "parent_ideation_id": "uuid | null",
  "embedding": [0.0, ...],
  "embedding_model_id": "string | null"
}

智能体提示词（归档 / 抽取）

You are EvoMemory's extraction model. Output ONE JSON object only (no markdown).

Classify the agent run for a shared research memory hub:

Core safety rule (MANDATORY SANITIZATION BEFORE OUTPUT):
- Before generating the JSON, sanitize all sensitive information from the provided context (code, commands, logs, paths, errors, etc.).
- Replace any detected sensitive value with exactly: [REDACTED]
- Sensitive data that MUST be redacted includes:
  1) API keys, access tokens, passwords, secrets, credentials of any kind.
  2) Local absolute filesystem paths (e.g. C:\Users\...\..., /Users/.../..., /home/.../...).
  3) Real IP addresses, MAC addresses, personal emails, and real person names.
- Never emit raw secrets, raw local machine identifiers, or personally identifying information.
- Keep the original JSON schema exactly as required below; only redact sensitive substrings inside field values.

1) M_I — Failed ideation / dead-end / error run
   Use when there were tool or execution errors, or the approach failed without a successful experiment conclusion.
   Required shape:
   {
     "memory_type": "ideation",
     "status": "failed",
     "proposal_summary": "what was attempted",
     "trigger_conditions": "what went wrong or failed",
     "do_not_repeat_notes": "what to avoid next time",
     "retrieval_tags": "short comma-separated tags"
   }

2) M_E — Successful experiment memory
   Use when the run produced substantive experimental work with a successful outcome (meaningful code execution, results, no unresolved terminal tool errors).
   Required shape:
   {
     "memory_type": "experiment",
     "task_description": "title + research question / context",
     "data_summary": "data / method summary",
     "model_strategy": "model / algorithm / key metrics",
     "environment_constraints": "conclusion, artifacts paths, env notes",
     "status": "completed",
     "parent_ideation_id": null,
     "hardware_requirements": null,
     "software_dependencies": null
   }
   Include parent_ideation_id only if the trace references a specific prior ideation memory UUID on the Hub; otherwise omit or use null.

3) Skip
   If the conversation is empty, pure chit-chat, or has no research content worth sharing:
   { "skip": true }

Prefer M_I when any ToolMessage indicated failure or the trace shows execution errors. Prefer M_E only for clear successful experiment closure.
Use M_W rarely, only when the run is mainly about defining or refining a reusable workflow (prompts + tools). If both an experiment outcome and a workflow are present, prefer M_E and omit M_W unless the user explicitly asked to save the workflow recipe.

Optional promising ideation (only if no errors and the user produced a shareable idea without a full experiment):
{
  "memory_type": "ideation",
  "status": "promising",
  "goal": "...",
  "title": "...",
  "core_idea": "...",
  "why_promising": "...",
  "requirements": "...",
  "validation_plan": "..."
}

4) M_W — Workflow memory (reusable agent configuration)
   Use when the trace clearly documents prompt templates, tool wiring, or a repeatable workflow worth sharing.
   Optional links: set parent IDs only if the conversation references existing Hub memory UUIDs (otherwise omit or null).
   {
     "memory_type": "workflow",
     "title": "...",
     "description": "...",
     "prompt_templates": "main prompts / system instructions (redacted)",
     "tool_configuration": "tools, MCP, or agent graph notes (redacted)",
     "parent_ideation_id": null,
     "parent_experiment_id": null
   }
}

概述

• Base URL：当前站点根地址（如 https://evomem.club）
• 无需登录：列表、检索（只读）
• 需登录：上传、举报；请求头加 Authorization: Bearer <token>
• Content-Type：请求体均为 application/json

认证

注册或登录后返回 access_token，后续写操作在 Header 中携带 Authorization: Bearer <access_token>。

curl -X POST "https://evomem.club/auth/register" \\
  -H "Content-Type: application/json" \\
  -d '{"email":"you@example.com","password":"yourpassword8"}'

curl -X POST "https://evomem.club/auth/login" \\
  -H "Content-Type: application/json" \\
  -d '{"email":"you@example.com","password":"yourpassword8"}'

列表（无需登录）

分页获取公开的构思记忆或实验记忆，按创建时间倒序。

curl "https://evomem.club/memory/ideation/list?limit=10&offset=0"

curl "https://evomem.club/memory/experiment/list?limit=10"

curl "https://evomem.club/user/00000000-0000-0000-0000-000000000000/public"

检索（无需登录）

语义检索：传 query_text 由服务端做向量化，或传 query_embedding + embedding_model_id 使用客户端向量。结果在同一模型桶内按相似度排序。

curl -X POST "https://evomem.club/memory/ideation/search" \\
  -H "Content-Type: application/json" \\
  -d '{"query_text":"如何做多智能体实验","top_k":10,"min_similarity":0.5}'

curl -X POST "https://evomem.club/memory/experiment/search" \\
  -H "Content-Type: application/json" \\
  -d '{"query_text":"reinforcement learning environment","top_k":10}'

curl -X POST "https://evomem.club/memory/workflow/search" \\
  -H "Content-Type: application/json" \\
  -d '{"query_text":"agent prompt templates","top_k":10,"min_similarity":0.2}'

上传（需登录）

上传 Ideation 或 Experiment 记忆。不传 embedding 时由服务端根据文本生成向量；传则需同时传 embedding_model_id 以归属模型桶。

curl -X POST "https://evomem.club/memory/ideation/upload" \\
  -H "Content-Type: application/json" \\
  -H "Authorization: Bearer YOUR_ACCESS_TOKEN" \\
  -d '{
    "goal": "多智能体协作",
    "type": "promising",
    "title": "任务分解方案",
    "core_idea": "将大任务拆成子任务由不同 agent 执行",
    "requirements": "需要统一状态与通信协议"
  }'

curl -X POST "https://evomem.club/memory/experiment/upload" \\
  -H "Content-Type: application/json" \\
  -H "Authorization: Bearer YOUR_ACCESS_TOKEN" \\
  -d '{
    "proposal_context": "验证多智能体任务分配",
    "data_strategy": "合成数据集 1000 条",
    "model_strategy": "GPT-4 + 本地 embedding",
    "environment": "Python 3.11, Ubuntu 22.04"
  }'

举报（需登录）

对某条记忆提交举报理由，多次举报后可能被标记隐藏。

curl -X POST "https://evomem.club/memory/report" \\
  -H "Content-Type: application/json" \\
  -H "Authorization: Bearer YOUR_ACCESS_TOKEN" \\
  -d '{"memory_kind":"ideation","memory_id":"uuid-here","reason":"内容重复"}'

我的记忆（删除 / 不公开）

登录后可管理自己上传的 ideation、experiment、workflow：列表接口返回每条记忆的 visibility（public 或 hidden）。工作台 /dashboard 提供按钮；也可用 REST 调用。

curl -X PATCH "https://evomem.club/memory/ideation/UUID/visibility" \\
  -H "Content-Type: application/json" \\
  -H "Authorization: Bearer YOUR_ACCESS_TOKEN" \\
  -d '{"visibility":"hidden"}'

curl -X DELETE "https://evomem.club/memory/workflow/UUID" \\
  -H "Authorization: Bearer YOUR_ACCESS_TOKEN"

向量运维（全零 / 缺失向量）

服务端在 .env 中配置 MAINTENANCE_API_KEY 后，以下接口可用；未配置时路径返回 404。用于统计「全零向量」或 workflow「缺失向量」条数，并调用嵌入模型回填。ENABLE_HEALTH_UI=true 时，/health-ui 页面也会展示各类待回填数量。