改进版 AI 长期记忆系统:检索与组织解耦,纯语义向量检索,四层记忆栈。兼容 Claude Code / Gemini CLI / 任意 MCP 客户端。
VecRecall has 23 tools with complete JSON schemas and descriptions in Chinese. However, critical gaps limit production readiness: (1) Parameter descriptions are minimal or absent for many fields (e.g., mp_add's 'wing', 'importance', 'session_id' lack guidance on format/constraints); (2) Output schemas are completely undocumented, no tool declares what fields it returns, forcing LLMs to guess; (3) Tool names use 'mp_' prefix (memory pool?) but lack clear verb-noun structure; (4) Error handling is absent, no guidance on retryability, user-fixable vs fatal errors, or recovery paths; (5) No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear read/write semantics. Schemas are present and mostly well-formed, but descriptions lack the specificity (format, range, constraints) needed for reliable LLM invocation. Chinese descriptions are clear but do not follow Arcade's 50-200 char guideline for LLM optimization.
添加一条记忆(逐字存储原文,向量化索引)
批量添加记忆
自动存档整段对话(Hooks 后台调用)
浏览某个话题下的记忆(UI 展示)
浏览某个 wing 下的记忆(UI 展示,不走向量检索)
构建四层记忆上下文(L0+L1+L2+L3),直接注入 AI prompt
读取某个 Agent 的日记(语义检索)
Agent 日记写入(每个 Agent 有独立 wing,互不干扰)
No output schemas documented. Tools return results but LLMs cannot predict field names, types, or structure. Forces agents to guess and parse unstructured responses.
Parameter descriptions lack format/constraint guidance. E.g., 'importance' (0-1 range), 'threshold' (0.55 default), 'limit' (max value?) are underdocumented. LLMs cannot validate inputs without explicit constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
导出某个 wing 的全部记忆为 JSON
模糊回忆:'我们之前聊过类似的东西吧?'——低阈值宽松语义匹配
按 ID 获取单条记忆(含原文和 ui_summary)
健康检查:验证向量库和 SQLite 连接
从 JSON 文件导入记忆
获取 L1 关键时刻(按 importance 排序的 top-15 记忆)
语义触发 L2 话题上下文(相似度 ≥ 阈值才返回)
L3 全量深度语义检索(最高召回,按需触发)
建立两条记忆之间的跨 wing 关联
列出所有话题标签——仅 UI 组织用
列出所有 wing(项目/人物)——仅 UI 组织用
语义搜索(纯向量,无结构过滤,高召回率)
更新 L0 身份层内容
系统状态:记忆总数、wing 数、跨关联数
更新一条记忆的重要性评分
No error handling guidance. Tools do not document retryability, user-fixable errors, or recovery paths. E.g., what happens if a memory_id is invalid? Is it retryable? Should the agent search first?
Tool names lack clear verb-noun structure. 'mp_add', 'mp_search', 'mp_build_context' are acceptable, but 'mp_l1_moments', 'mp_l2_context', 'mp_l3_deep' are cryptic. LLMs cannot infer intent from 'L1/L2/L3' abbreviations without domain knowledge.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Write tools (mp_add, mp_update_importance, mp_link, mp_diary_write, mp_archive_session, mp_import_json, mp_set_identity) should be marked destructiveHint=true for safety. Read tools should be marked readOnlyHint=true.