-
Notifications
You must be signed in to change notification settings - Fork 53
When to Use
AgentRecall is not "always on" overhead. It's a tool with a clear break-even point.
AgentRecall 不是"始终开启"的开销。它是一个有明确盈亏平衡点的工具。
Layer 1 (5 seconds): It makes your AI agent remember what happened last session.
第一层(5 秒): 让你的 AI agent 记住上次会话发生了什么。
Layer 2 (30 seconds): Every time you correct your agent — "no, not that version", "ask me first" — that correction is stored permanently and recalled before the agent makes the same mistake again. After 10 sessions, your agent understands your priorities, your communication style, your non-negotiables.
第二层(30 秒): 每次你纠正 agent——"不,不是那个版本"、"先问我"——这个纠正被永久存储,并在 agent 再犯同样错误之前被召回。10 次会话后,你的 agent 理解你的优先级、你的沟通风格、你的不可妥协项。
Layer 3 (2 minutes): The Intelligent Distance Protocol. The structural gap between human thinking and AI action can't be closed — but it can be navigated better every session. Corrections are training data. The 200-line awareness cap forces quality over quantity. Cross-project insights mean lessons learned once apply everywhere. See Intelligent Distance for the full protocol.
第三层(2 分钟): 智能距离协议。人类思维和 AI 行动之间的结构性鸿沟无法消除——但可以在每次会话中更好地穿越。纠正就是训练数据。200 行感知上限强制质量优于数量。跨项目洞察意味着学到一次的经验到处适用。完整协议参见 智能距离。
Default: USE IT. Most projects are long-term and benefit from memory. Memory compounds — a small overhead today saves large context-rebuilding costs across future sessions.
默认:使用它。 大多数项目是长期的,记忆对它们有益。记忆是复利的——今天的小开销,能节省未来会话中大量的上下文重建成本。
Skip only when the task is truly single-session throwaway work.
仅在 任务确实是一次性的临时工作时跳过。
| Situation | Why | 为什么 |
|---|---|---|
| Multi-session project (3+ sessions expected) | Memory compounds across sessions | 记忆跨会话复利 |
| Resuming work from a previous session | Cold start loads context in ~200 tokens | 冷启动用约 200 tokens 加载上下文 |
| Non-obvious decisions being made | Future agents need to know WHY, not just WHAT | 未来的 agent 需要知道「为什么」而不仅是「是什么」 |
| Multiple people or agents touch the same project | Shared memory prevents repeated mistakes | 共享记忆防止重复犯错 |
| Cross-project work | Insights from Project A surface in Project B | 项目 A 的洞察在项目 B 中浮现 |
| Situation | Why | 为什么 |
|---|---|---|
| Pure Q&A session | No project context to save | 没有需要保存的项目上下文 |
| Trivial one-off script | Won't be revisited | 不会再用 |
| Quick fix with no decisions | Nothing worth recalling | 没有值得回忆的内容 |
We ran a controlled experiment (2026-04-10) comparing token usage with and without AgentRecall on a simple CLI task (CSV-to-JSON converter):
我们进行了一个对照实验(2026-04-10),比较在简单 CLI 任务(CSV 转 JSON 工具)上使用和不使用 AgentRecall 的 token 用量:
| Metric / 指标 | Without AR / 无 AR | With AR / 有 AR | Delta / 差异 |
|---|---|---|---|
| Total tool calls / 总工具调用 | 9 | 17 | +8 (+89%) |
| Functional tool calls / 功能性调用 | 9 | 9 | 0 |
| AR tool calls / AR 工具调用 | 0 | 8 | +8 |
| Est. token overhead / 预估 token 开销 | 0 | ~2,300 | +~30% |
| Corrections needed / 需要修正次数 | 0 | 0 | 0 |
| Rework count / 返工次数 | 0 | 0 | 0 |
Result for simple task: pure overhead. No insight matched. No prior context was useful.
简单任务的结果:纯开销。 没有匹配到任何洞察。没有先前上下文有用。
But this is the exception, not the rule. 但这是例外,不是常态。
For a multi-session project, the math flips:
对于多会话项目,算术反转:
| Scenario / 场景 | AR overhead / AR 开销 | Context rebuild cost without AR / 无 AR 上下文重建成本 | Net / 净效果 |
|---|---|---|---|
| 1 session, simple task / 1 次会话,简单任务 | ~2,300 tokens | 0 | -2,300 (waste / 浪费) |
| 3 sessions, medium project / 3 次会话,中等项目 | ~6,900 tokens | ~5,000-10,000 tokens re-explaining | Break-even / 持平 |
| 10 sessions, complex project / 10 次会话,复杂项目 | ~23,000 tokens | ~50,000-100,000 tokens lost context | +27,000-77,000 saved / 节省 |
AgentRecall usage is a dynamic equilibrium, not a binary switch.
AgentRecall 的使用是一个动态平衡,不是一个二元开关。
The right question is not "should I use it?" but "will a future session benefit from today's context?"
正确的问题不是「我应该使用它吗?」而是「未来的会话会从今天的上下文中受益吗?」
- Cost is immediate — tokens spent now on tool calls / 成本是即时的——当下花在工具调用上的 tokens
- Value is deferred — future sessions benefit from today's writes / 价值是延迟的——未来的会话从今天的写入中受益
- Value compounds — each insight strengthens or replaces, so 100 sessions later, memory is still 200 lines but each line carries more weight / 价值是复利的——每条洞察强化或替换旧的,所以 100 次会话后,记忆仍然是 200 行,但每一行都承载更多分量
For most real work, the long-term benefit far outweighs the per-session cost.
对于大多数实际工作,长期收益远超单次会话成本。
- Getting Started — install and first session / 安装和第一次使用
- Core Concepts — how the memory layers work / 记忆层如何工作
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