pasm-mcp-server
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| PASM_EMBED_KEY | No | 可选:外部 embedding 接口的密钥 | |
| PASM_EMBED_URL | No | 可选:外部 embedding 接口的 URL | |
| PASM_EMBED_MODEL | No | 可选:外部 embedding 模型的名称 | |
| PASM_MCP_PERSONA | No | 默认人格(JSON 字符串) | |
| PASM_SKILLS_PATH | No | 本地开发逃生口:指向 pasm-skills 仓库根目录 | |
| PASM_MCP_PERSIST_DIR | No | 状态落盘根目录 | ~/.pasm-mcp/ |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
PASM 认知内核:提供长期记忆、情绪状态与动作倾向。典型用法是先 pasm_context 取认知上下文,再据此回复;有值得记住的事就 pasm_observe 写进去;被夸/被凶就用 pasm_feedback 塑形。
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| pasm_contextA | 取 PASM 认知上下文(只读,不写记忆)。返回:与当前话题相关的长期记忆、当前情绪值、人格、可用动作池。这是把 PASM 接进大模型提示词的首选工具 —— 先调它,再把 recalled 里的内容当作'你确实记得的事'写进回复。 |
| pasm_recallC | 在 PASM 长期记忆里检索与 query 相关的经历,返回标题/摘要/标签/重要度。 |
| pasm_semanticA | 语义检索(认知层):换一种说法也能命中,并解释每条为什么被召回。适合当 pasm_recall 检索不到时的第二步 —— 它走向量 + 中文同义扩展,并按'记忆保留度'加权(很久没被唤起的记忆会自然降权)。 |
| pasm_focusA | 焦点栈:查看或更新'现在在聊什么'。长对话里先压入当前话题,后续检索会自动给相关内容加权,避免跑题。不传 topic 就是查看。 |
| pasm_consolidateA | 记忆巩固(睡眠回放):把高度相似的重复经历蒸馏成一条要点,并把源条目标记为已合并,避免记忆池被重复项撑爆。默认只给建议(apply=false),确认后传 apply=true 执行。 |
| pasm_observeA | 把一件事写进 PASM 长期记忆。值得记住的用户信息、承诺、偏好、共同经历都应该写进来。salience 越高越不容易被淘汰(1-5,重要的事用 4-5)。 |
| pasm_feelA | 向 PASM 报告一个带情绪效价的事件,驱动情绪状态演化(valence: -1 负面 ~ +1 正面)。 |
| pasm_actB | 让 PASM 按「人格基线 + 反馈学到的偏好」从动作池里选一个动作。不传 candidates 就用智能体自己的动作池。 |
| pasm_feedbackA | 用户对 PASM 行为的反馈,会塑形后续动作选择。kind: praise(夸奖)/scold(责怪)/poke(调侃)/hug(安慰)/ignore(无视)。必须尽量带 action(被评价的那个动作名),否则偏好会集中在某一个动作上。 |
| pasm_chatA | 走一次完整认知回路:检索记忆 → 渲染回复 → 把这次对话写进记忆。回复是模板渲染、非 LLM 生成,用于快速验证或纯离线场景;要自然语言效果请用 pasm_context 配合你自己的模型。 |
| pasm_personaA | 查看或更新人格。不传 persona 就是查看。传入为合并更新(未提及的键保持原值)。常用键:name/role/tone/temper(0-1 主动性)/energy(0-1 活跃度)/play(0-1 俏皮度)/actions(动作池数组)。 |
| pasm_statusB | 查看 PASM 状态快照:档位(light/core/bionic)、记忆条数、情绪值、交互次数、落盘目录。 |
| pasm_saveA | 把 PASM 状态立即落盘(不传 agent_id 则保存全部已加载的智能体)。 |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 13 tools
Most tools have clearly distinct roles, and the descriptions actively sequence the three retrieval paths (pasm_context first, pasm_recall as keyword fallback, pasm_semantic as the semantic step), which resolves most of the overlap. Residual ambiguity remains between pasm_chat (which internally does retrieve+render+write) and the context/observe pair, and pasm_persona/pasm_focus both use the 'omit arg = read, pass arg = write' dual-mode pattern, which is easy to misselect.
Every tool follows the identical pasm_<lowercase_verb_or_noun> snake_case pattern with no exceptions or style drift. The single shared prefix also makes the tool family instantly recognizable.
13 tools sit squarely in the well-scoped range and each maps to a distinct cognitive sub-function (persona, status, save, retrieve, write, focus, consolidate, emotion, action, feedback, chat). Nothing appears redundant enough to drop.
The memory lifecycle covers write (observe), retrieve (recall/semantic/context), and consolidate, but there is no delete/forget operation, no way to edit or fetch a single memory by ID, and no list/browse with pagination — an agent that writes a wrong or stale memory has no direct correction path. Consolidation offers only partial mitigation.