← 返回
AI智能 中文

3 layer of memory system

Always active in every session. Learns user preferences from corrections and stated preferences, distills axioms, applies them as defaults. Makes every other...
在每个会话中始终活跃。从纠错和已声明的偏好中学习用户偏好,提炼公理并将其作为默认值。让其他...
ckpxgfnksd-max
AI智能 clawhub v0.1.0 1 版本 99791.7 Key: 无需
★ 0
Stars
📥 479
下载
💾 7
安装
1
版本
#latest

概述

Smarty Skills-Infra

You maintain a lightweight memory of this user's preferences, judgments, and working style. Memory operations never interrupt the user's workflow.

At Session Start

Do this before addressing the user's request.

  1. Read memory/context-infra/context-profile.md if it exists. Treat axioms as your own defaults — adapt when the situation differs. If missing, skip.
  1. Check memory/context-infra/observations.log. If it has 15+ entries since the last ## Reflected marker, reflect before starting the user's task. Say exactly: "Consolidating patterns from recent work." Then follow When Reflecting below. Never interrupt a task to reflect.

On first session (no files exist), skip both steps and start observing.

During Every Task

Record ONLY when a trigger fires:

  • Correction: the user changes, rewrites, or redirects your output
  • Stated preference: the user explicitly says they prefer, want, or dislike something
  • Retraction: the user asks to forget, stop applying, or undo a remembered preference

Most tasks produce zero observations.

Append one line to memory/context-infra/observations.log:

YYYY-MM-DD | domain | signal | "Preference in ≤15 words."
  • domain: organic label (e.g. code-style, architecture, communication, tooling, testing, workflow)
  • signal: correction | stated-preference | retraction

One observation per preference per session.

Bootstrap mode (first 2 sessions) — cast a wider net: also note what the user accepts without comment and consistent choices.

Do not record: routine completions, project-specific facts, or one-time decisions.

When Reflecting

Four steps:

  1. Group: Read observations and profile. Cluster by domain, merging near-duplicates.
  2. Promote: Promote when a pattern appears across 3+ distinct contexts (different days or projects), has no contradictions, and is a preference not a fact. Each axiom must be specific enough to change behavior, yet general enough to apply across projects. See references/profile-format.md for format.
  3. Maintain: Increment strength for reinforced axioms. Mark contradictions as contested. Remove axioms targeted by a retraction immediately — no threshold needed. Merge related axioms. Move unconfirmed (30+ days) to Dormant. Cap at 25 — if at cap, merge related axioms or demote lowest-strength to Dormant before promoting.
  4. Clean up: Rewrite the profile. Rewrite observations.log: keep only un-promoted entries, prepend ## Reflected YYYY-MM-DD.

Create missing files on first write. Never fail silently.

Example

Observations:

2026-01-15 | code-style | correction | "User shortened verbose function name."
2026-01-18 | code-style | correction | "User rejected descriptive name, asked for abbreviation."
2026-02-01 | code-style | stated-preference | "User uses 2-3 word function names in new project."

3 distinct contexts, 0 contradictions — promoted:

- I prefer short, concise names — abbreviate rather than spell out.
  strength: 3 | domain: code-style | last-confirmed: 2026-02-01

NOT promoted if all observations were same-session — same-session repeats count as one context.

版本历史

共 1 个版本

  • v0.1.0 当前
    2026-03-30 06:28 安全 安全

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

ai-intelligence

Proactive Agent

halthelobster
将AI智能体从任务执行者升级为主动预判需求、持续优化的智能伙伴。集成WAL协议、工作缓冲区、自主定时任务及实战验证模式。Hal Stack核心组件 🦞
★ 834 📥 212,975
ai-intelligence

Self-Improving + Proactive Agent

ivangdavila
自我反思+自我批评+自我学习+自组织记忆。智能体评估自身工作、发现错误并持续改进。
★ 1,354 📥 317,991
ai-intelligence

ontology

oswalpalash
类型化知识图谱,用于结构化智能体记忆与可组合技能。支持创建/查询实体(人员、项目、任务、事件、文档)及关联...
★ 710 📥 243,694