-
Notifications
You must be signed in to change notification settings - Fork 53
FAQ
No. Everything is stored as local markdown files in ~/.agent-recall/. Zero cloud, zero telemetry, zero API keys. Browse your data in Obsidian, grep it in the terminal, version it with git.
不会。 所有数据都以本地 markdown 文件存储在 ~/.agent-recall/ 中。零云服务、零遥测、零 API key。在 Obsidian 中浏览，在终端中 grep，用 git 版本管理。
Any MCP-compatible agent: Claude Code, Cursor, VS Code (Copilot), Windsurf, Codex. Plus SDK for any JS/TS framework (LangChain, CrewAI, Vercel AI SDK) and CLI for terminal workflows.
任何兼容 MCP 的 agent：Claude Code、Cursor、VS Code (Copilot)、Windsurf、Codex。另外 SDK 支持任何 JS/TS 框架（LangChain、CrewAI、Vercel AI SDK），CLI 支持终端工作流。
No. Start with 5: recall_insight, palace_walk, journal_capture, alignment_check, awareness_update. These cover 90% of daily use. See MCP Tools Reference for the priority guide.
不需要。 从 5 个开始：recall_insight、palace_walk、journal_capture、alignment_check、awareness_update。覆盖 90% 的日常使用。参见 MCP 工具参考 的优先级指南。
Skip it for truly throwaway single-session tasks: quick Q&A, trivial one-off scripts, simple fixes. For everything else, use it — memory compounds over time. See When to Use for the full decision guide.
对于真正一次性的单会话任务跳过它：快速问答、简单的一次性脚本、简单修复。其他情况都使用——记忆随时间复利。完整决策指南参见 何时使用。
Yes. The palace builds up over sessions. After your first /arsave, you'll have journal entries and initial palace rooms. After 3-5 sessions, the awareness system starts compounding insights.
是的。 记忆宫殿随会话逐渐积累。第一次 /arsave 后，你会有日志条目和初始宫殿房间。3-5 次会话后，感知系统开始复利洞察。
Measured on a simple CLI task: ~2,300 tokens (~30% overhead) per session from MCP tool calls. For simple single-session tasks, this is pure overhead. For multi-session projects, the cold-start recall saves far more than it costs. Break-even point: ~3 sessions.
在简单 CLI 任务上测量：每次会话来自 MCP 工具调用约 2,300 tokens（约 30% 开销）。对于简单的单会话任务，这是纯开销。对于多会话项目，冷启动回忆节省的远超成本。盈亏平衡点：约 3 次会话。
Yes. Memory is file-based. Any agent pointing to the same ~/.agent-recall/ directory reads the same data. This is how /arstart works — a fresh agent instance loads yesterday's decisions written by a different agent.
可以。 记忆基于文件。任何指向相同 ~/.agent-recall/ 目录的 agent 读取相同数据。这就是 /arstart 的工作方式——一个新的 agent 实例加载昨天由不同 agent 写入的决策。
This means no stored insight matches your current task context. This is normal for new projects or when working in a new domain. Insights build up over time through awareness_update.
这意味着没有存储的洞察匹配你当前的任务上下文。这对于新项目或在新领域工作时是正常的。洞察通过 awareness_update 随时间积累。
Check the depth parameter:
-
identityreturns ~50 tokens (just the project name) -
activereturns ~200 tokens (top rooms) - Use
relevantwith afocusparameter for targeted loading - Use
fullonly when you need everything
检查 depth 参数：
-
identity返回约 50 tokens（只有项目名） -
active返回约 200 tokens（顶部房间） - 使用
relevant加focus参数进行定向加载 - 只在需要全部内容时使用
full
Use journal_rollup to condense old journals into weekly summaries:
使用 journal_rollup 将旧日志压缩为周报摘要：
# Preview what would be rolled up
ar rollup --dry-run
# Roll up entries older than 14 days
ar rollup --min-age-days 14This is by design. The 200-line cap forces merge-or-replace. When a new insight is added, it either merges with an existing one (strengthening it) or replaces the lowest-salience insight. This is what makes memory compound instead of accumulate.
这是设计如此。 200 行上限强制合并或替换。当添加新洞察时，它要么与现有的合并（强化它），要么替换最低显著性的洞察。这就是让记忆复利而不是线性累积的机制。
See the Migration Guide for breaking changes and upgrade steps.
破坏性变更和升级步骤参见 迁移指南。
- Getting Started — install and first session / 安装和第一次使用
- When to Use — decision guide / 决策指南
- MCP Tools Reference — all tools explained / 全部工具详解
AgentRecall Wiki
Get Started | 快速上手
Concepts | 核心概念
Reference | 参考手册
Guides | 使用指南
Other | 其他
Dev Log | 开发日志