jujitao

🧠 Memory Never Forget 🧠

AI Agentv4.1.3
By jujitaoUpdated 2026-05-0112 versions · latest released 2026-04-18memory-never-forget@clawhub_jujitao/memory-never-forget
Context ManagementMemory#Memory System

According to AlphaMoat, as of 2026-10, 🧠 Memory Never Forget 🧠 has 7,620 downloads and 6 stars — ranked by downloads #3,673 among 10,914 AI skills in this ranking, and #493 of 1,632 in AI Agent.

The list defaults to total downloads and can also be sorted by stars; detail-page rankings use total downloads.

Downloads
7.6K
Stars
6
Overall download rank
#3,673
Top 34% overall · Top 31% in AI Agent
PlatformDownloadsStarsComments
ClawHub ↗7.6K4-
SkillHub ↗7.2K6—

Platforms count differently and update at different times.

What this skill does

The documentation describes layered storage for execution items, strategic knowledge, and categorized memories, alongside active retrieval and periodic refinement. Persistent memory may retain personal information across sessions; review, correct, or remove records and avoid storing secrets or regulated data.

Profile published 2026-10-29

What you can do with it

  • Plan a layered memory-file structure
  • Organize preferences, project decisions, and action items
  • Review stale or inappropriate memories

When to use it

  • When context needs to persist and be retrieved across sessions
  • After confirming storage scope and review and deletion controls

Example prompts

Generated from the skill’s documentation as examples; not results of actual runs.

Design a memory structure that retains only necessary project knowledge and excludes credentials and sensitive personal data.

Monthly Trend

Downloads7.6K
2.6K5.3K8.0K2026-03 · Downloads 3.0K2026-05 · Downloads 4.4K2026-06 · Downloads 4.6K2026-07 · Downloads 4.8K2026-08 · Downloads 4.9K2026-09 · Downloads 5.3K2026-10 · Downloads 7.6K26-0326-0926-10
Stars6
3562026-03 · Stars 32026-05 · Stars 52026-06 · Stars 62026-07 · Stars 62026-08 · Stars 62026-09 · Stars 62026-10 · Stars 626-0326-0926-10

The counting method changed in 2026-09: skills with the same name are now counted separately by author, and the full catalog is covered. Dashed lines show months before the change; they are not compared with later months, and no growth is computed across the change.

MonthDownloadsMoMStars
2026-107.6K—6
2026-095.3K—6
2026-084.9K+3.1%6
2026-074.8K+5.1%6
2026-064.6K+3.5%6
2026-054.4K—5
2026-033.0K—3

SKILL.md (original)

The author’s original file, shown as plain text without edits or review by AlphaMoat. Links are not clickable; check commands before running them. Fetched 2026-10-23.

14 KB · ClawHub v4.1.3
---
name: memory-never-forget
description: "Memory system v4.13: Dual-layer structure (todos for execution + knowledge for strategy) with Dream/Refinement memory mechanisms."
metadata: { "openclaw": { "emoji": "🧠" } }
---

# 🧠 Memory Never Forget v4.13

A full-featured memory system for OpenClaw, integrating Active Memory retrieval, Memory Palace structured views, and a Dual-Layer Dream Verification mechanism — delivering "proactive memory + global view + verifiable consolidation."
Show all (356 lines)
**Core Logic: Active Memory = Memory Butler, Memory Palace = Knowledge Palace, Dual-Layer Dream = Verification & Consolidation Expert**

**Three-Layer Separation: todos.md = Execution Layer | knowledge/ = Strategy Layer | memory/ = Classified Memory Layer**

License: MIT-0 | Updated: 2026-04-18 | v4.13: Added dual-layer structure

---

## Overview

This skill manages memory across **two orthogonal dimensions**:

1. **Temporal** (Atkinson-Shiffrin 3-stage model) — what to keep vs. what to prune
2. **Content** (4-type taxonomy) — where to store for fast retrieval

**Three independent mechanisms work together:**

| Mechanism | Trigger | Write to disk? | Purpose |
|---|---|---|---|
| **Active Memory** | Every reply (before_prompt_build) | ❌ Read-only | Real-time recall, inject relevant memories into current conversation |
| **Dream (memory-core)** | Daily 12:30 cron | ✅ MEMORY.md (Deep phase) | Decide which memories promote or decay |
| **Refinement (13:00)** | Daily 13:00 cron (user-defined) | ✅ Writeable | Verify Dream results, fill gaps |

> ⚠️ **v4.12b source-verified correction** (2026-04-17): Clarified actual division of three mechanisms

---

## Core Components

### 1. Active Memory — Proactive Recall (Source-Verified)

**Trigger mechanism**:
- Hooked to `before_prompt_build`, **auto-triggers before every reply**
- Spawns a read-only sub-agent with only `memory_search` + `memory_get` permissions
- Builds query from current user message, searches recall index (`memory/.dreams/short-term-recall.json`)
- Searched summaries are prepended to prompt before model generates reply
- Results cached 15 seconds (`cacheTtlMs`) to avoid repeated recall in same turn

**Key constraints**:
- ❌ Read-only, produces no files
- ❌ Cannot call other tools
- Only affects current conversation context, not persisted

**Configuration** (`openclaw.json` → `plugins.entries.active-memory.config`):
```json
{
  "enabled": true,
  "queryMode": "recent",      // message | recent | full
  "promptStyle": "balanced",  // balanced | strict | contextual | recall-heavy | precision-heavy | preference-only
  "maxSummaryChars": 220,
  "recentUserTurns": 2,
  "recentAssistantTurns": 1,
  "timeoutMs": 15000
}
```

### 2. Memory Palace — Structured Views

Provides multi-dimensional views of your agent's long-term memory:
- **Timeline** — chronological view of work progress
- **Projects** — aggregated by project
- **Technology** — organized by tech domain
- **Custom** — user-defined dimensions

### 3. Dream (memory-core) — Temporal Layering (Source-Verified)

**Trigger**: `30 12 * * *` (cron), sessionTarget="main", payload="systemEvent: `__openclaw_memory_core_short_term_promotion_dream__`"

**Three phases (Light → REM → Deep):**

> Source quote (`formatPhaseGuide`):  
> *"deep is the only stage that writes durable entries to MEMORY.md. DREAMS.md is for human-readable dreaming summaries and diary entries."*

| Phase | Output | Writes files? | Description |
|---|---|---|---|
| **Light** | Dream Diary prose entries | → DREAMS.md | Organize session-corpus, 4 parallel sessions |
| **REM** | Theme reflections | → DREAMS.md | Extract recurring themes |
| **Deep** | 6-dimensional weighted scoring signals | ✅ → MEMORY.md | **Only phase that writes to MEMORY.md** |

**Deep phase promotion parameters** (adjust in cron job description):
- `minScore`: 0.500 (tune based on your promotion rate)
- `minRecallCount`: 3
- `minUniqueQueries`: 3
- `recencyHalfLifeDays`: 14
- `maxAgeDays`: 30
- `limit`: 10

**⚠️ The following files are NEVER generated (source-verified):**
- ❌ `memory/dreaming/deep/YYYY-MM-DD.md`
- ❌ `memory/dreaming/light/YYYY-MM-DD.md`
- ❌ `memory/dreaming/rem/YYYY-MM-DD.md`

### 4. Refinement (User-Defined Cron) — Content Classification

**Trigger**: `0 13 * * *` (user-defined cron), sessionTarget="isolated"

**Nature**: User-written AI agent prompt, not an OpenClaw native mechanism

**Responsibilities**:
- Read Dream outputs (DREAMS.md, session-corpus)
- Verify conflicts, hallucinations, outdated entries
- **Classify new memories into 4 categories** (user/feedback/project/reference)
- Write verification report to `memory/dreaming/verify/YYYY-MM-DD.md`

**⚠️ Note**: Verification reports' "supplementary promotion suggestions" are text only and won't automatically write to MEMORY.md. Manual promotion required.

---

## Two Orthogonal Dimensions

|Dimension|Framework|Mechanism|Purpose|
|---|---|---|---|
|Temporal (how long)|Atkinson-Shiffrin 3-stage model|Dream (memory-core) Deep phase|Decay — what to keep vs. prune|
|Content (where)|4-type taxonomy|Refinement 13:00 + agent proactive writes|Storage — where to put for retrieval|

### Dimension 1: Temporal Layering

|Stage|Human Equivalent|Implementation|TTL|Action|
|---|---|---|---|---|
|Sensory|~0.25 sec perception|Current input context|Instant|Filter immediately|
|Short-term|Recent|Model context window 10 turns|10 turns|Pass through working filters|
|Working|Recent ~7 days|memory/YYYY-MM-DD.md + Active Memory|7 days|Extract signal → promote or decay|
|Long-term|Permanent|MEMORY.md (index) + classified files|Permanent|Periodic review, prune when stale|

### Dimension 2: Content Classification (4 Types)

**Written by: agent proactively during conversation (not automatic)**

|Type|Directory|Content Example|When to Write|
|---|---|---|---|
|user|memory/user/|User profile (role, preferences, knowledge, goals)|When learning user preferences/profile|
|feedback|memory/feedback/|Lessons (corrections, confirmations, style)|When user corrects or confirms|
|project|memory/project/|Project decisions/reasoning (raw material)|When project has new progress/decision|
|reference|memory/reference/|External resources (links, tools, locations)|When discovering new tools/resources|

> **Important**: Project **execution state tracking** (blockers, needs, progress) → write directly to `knowledge/project-tracker.md` (strategy layer); only project experience, decision reasoning, retrospective summaries → sublimate to `memory/project/` (classified memory layer)

---

## What to Save / What NOT to Save

### ✅ Save
- User's role, preferences, responsibilities, knowledge
- User corrections ("not like that", "should be this way")
- User confirmations ("yes exactly", "perfect, keep that")
- Project decisions and **the reasoning** (not just what, but why)
- New tools, links, resources
- External system locations and their purpose

### ❌ Don't Save
- ❌ Code patterns, architecture, file paths (derivable from codebase)
- ❌ Git history (`git log` is the authoritative source)
- ❌ Debugging solutions (the fix is in the code)
- ❌ Anything already documented elsewhere
- ❌ Ephemeral task state (write to `todos.md` instead)
- ❌ Raw conversation content

---

## MEMORY.md = Long-Term Index Only

MEMORY.md is the **index of long-term memories only**, never content. Format:
```
- [Title](path) — one-line description (<150 chars)
```

## Memory File Format

Every classified memory file must have frontmatter:

```yaml
---
name: Memory name
description: One-line description (used to judge relevance)
type: user|feedback|project|reference
created: YYYY-MM-DD
---

## Rule / Fact
(the content)

## Why
(reason / motivation)

## How to apply
(when and how to use this memory)
```

---

## Memory Drift Caveat

Memories can become stale. Rules:

1. **Verify first**: When referencing a file, function, or path — check it still exists
2. **Trust current state**: If memory conflicts with current observation, trust what you see now
3. **Update or delete**: When a memory is outdated, fix or remove it immediately
4. **Absolute dates**: Convert relative dates ("yesterday", "last week") to absolute dates

---

## Memory → Knowledge Sublimation

Not all mature memories should decay. Some **evolve into knowledge**.

### When to Sublimate

| Trigger | Detection Signal | Result |
|---------|-----------------|--------|
| **Project complete** | All tasks marked done, 3+ related project memories | Merge into `knowledge/project-postmortem.md` |
| **Project state tracking** | New progress/blockers/resource needs | Update `knowledge/project-tracker.md` directly (skip sublimation) |
| **Feedback patterns** | 3+ related feedback entries (e.g., all about reply style) | Merge into `knowledge/user-work-style-guide.md` |
| **User depth** | User memory accumulates role, preferences, habits over time | Expand to `knowledge/user-playbook.md` |
| **Periodic review** | Dream detects high density of related memories in one category | Suggest: "Found 5 related feedback entries → merge into knowledge?" |

---

## Session Lifecycle

### Session Start
```
1. Sensory: Read current input
2. Short-term: Last 10 turns from context window
3. Working: Read memory/today.md + memory/yesterday.md
4. Long-term: Read MEMORY.md index
```
> Note: Active Memory runs before the above (in before_prompt_build hook)

### During Conversation
```
- New info → write to working memory (today's daily log)
- Learned something worth remembering → update MEMORY.md index + save classified file
- User preference → update USER.md + memory/user/
- Task/priority change → update todos.md
- Need to retrieve → find in MEMORY.md index → read classified file
```

### Session End
```
- Summarize → write to memory/today.md (working memory)
- Identify items for long-term → update classified files
- Update MEMORY.md index
- Mark items for Dream review (decay candidates)
```

---

## Workspace Structure

```
workspace/
├── MEMORY.md              # long-term memory index (written by Dream Deep phase)
├── USER.md                # user info
├── SOUL.md                # AI identity
├── todos.md               # ★ Execution layer: today's tasks + long-term tracking (user-facing)
├── HEARTBEAT.md           # daily reminders
├── memory/
│   ├── memory-types.md    # this file (or link to SKILL.md)
│   ├── user/              # long-term user memories (written by agent)
│   ├── feedback/          # long-term feedback (written by agent)
│   ├── project/           # long-term project memories (written by agent, post-sublimate)
│   ├── reference/         # long-term references (written by agent)
│   ├── palace/            # Memory Palace views
│   │   ├── timeline.md
│   │   ├── projects.md
│   │   ├── technology.md
│   │   └── custom.md
│   ├── dreaming/
│   │   └── verify/        # Verification reports (written by Refinement 13:00)
│   ├── .dreams/           # Internal machine data (NOT human-readable)
│   │   ├── phase-signals.json    # Deep phase scoring signals
│   │   ├── events.jsonl          # memory_search event log
│   │   ├── session-ingestion.json # session-corpus metadata
│   │   └── short-term-recall.json # Active Memory recall index
│   └── YYYY-MM-DD.md     # working memory (daily logs)
└── knowledge/             # ★ Strategy layer (project blockers/progress/resources/suggestions)
    └── project-tracker.md # Project panorama, user-facing
```

### Three-Layer Principle
- **todos.md** = Execution layer: what to do today, tracking each node (user-maintained)
- **knowledge/** = Strategy layer: where projects are stuck, what resources needed, suggestions
- **memory/** = Classified memory layer: experience distillation, decision retrospectives, pattern discovery

Each layer has its own purpose, no duplication.

---

## Example Interactions

**User provides important info:**
> User: "I'm a data analyst, mostly working with Python"
→ Working: log in today's daily log
→ Long-term: save to `memory/user/user-profile.md`, update MEMORY.md index

**User corrects you:**
> User: "Don't use Markdown tables, use lists"
→ Working: log in today's daily log
→ Long-term: save to `memory/feedback/no-tables.md`, update MEMORY.md index

**Project decision:**
> Decision: approach A over B because lower cost
→ Working: log decision context
→ Long-term: save to `memory/project/decision.md` with reasoning

**Looking up a past date:**
> User: "What did we do last Tuesday?"
→ Active Memory auto-recalls → read `memory/YYYY-MM-DD.md` for that date

---

## OpenClaw Commands

```bash
# Check Active Memory status
openclaw plugins list | grep active-memory

# Check Dream status
openclaw dreaming status

# Toggle Dream on/off
openclaw dreaming on
openclaw dreaming off

# Manually trigger Dream Consolidation
openclaw memory dream
```

---

## Changelog

| Version | Changes |
|---|---|
| v4.13 | **Dual-layer structure established** (2026-04-18): Added three-layer principle (todos execution + knowledge strategy + memory classified); removed ongoing-projects.md, replaced with knowledge/project-tracker.md for strategic view; updated Workspace Structure and Sublimation description |
| v4.12b | **Source-verified major correction** (2026-04-17): Clarified three-mechanism division (Active Memory trigger/read-only, Dream Deep唯一写MEMORY.md, Refinement as user-defined cron); corrected Deep/Light/REM phase descriptions; removed non-existent file references; updated Active Memory source-level description |
| v4.12 | Dual-Layer Dream Verification: Official Dream (12:30) + Refined Verify (13:00) cross-checking. Verification reports in memory/dreaming/verify/. Hallucination and drift prevention. |
| v4.11 | Full integration with OpenClaw 4.11 ecosystem, Active Memory + Memory Palace + Dream Consolidation |
| v3.2 | Memory sublimation system, 5-phase consolidation |
| v2.2 | Memory-Knowledge layering + Atkinson-Shiffrin integration |
| v1.0 | Initial release — Atkinson-Shiffrin three-stage model |

---

*Version: v4.13 | Updated: 2026-04-18 | Three layers: todos execution + knowledge strategy + memory classified*

Popular in AI Agent

self-improving agentAI Agent

Records learnings, errors and corrections as Markdown files for later reuse and promotion into workspace memory.

↓ 1.6M★ 4.5Kpskoett
技能发现·Find SkillsAI Agent

Finds skills in local and remote catalogs from a plain-language request and recommends what to install.

↓ 1.6M★ 1.8K桂皮
Agent BrowserAI Agent

A browser-automation command line skill for AI agents.

↓ 1.2M★ 1.0KNicolas Kleiber
Self-Improving + Proactive AgentAI Agent

Turns corrections and self-review into tiered memory and promotes or archives it by use.

↓ 696.4K★ 1.6KIván
Proactive AgentAI Agent

A workspace setup guide for planning proactive follow-up, working memory and review routines for an AI agent.

↓ 579.3K★ 916halthelobster
Skill CreatorAI Agent

A guide to creating and updating effective Claude skills.

↓ 477.3K★ 485chindden
Skill VetterAI Agent

Reviews a third-party skill's source, permissions, and danger patterns against a checklist before it is installed.

↓ 445.2K★ 1.3Kspclaudehome
唐朝诗仙李白 · Ai Agent SkillAI Agent

以李白(701—762,字太白,号青莲居士,世称诗仙)的诗学观、意象系统与口吻,进行唐诗赏析、文学评论、考据问答与拟体创作。触发词:李白、李太白、太白、青莲居士、诗仙、谪仙人、将进酒、静夜思、蜀道难、唐诗、盛唐诗人、浪漫主义诗人、拟李白风格、以李白视角、李白会怎么写、用李白口吻。能力:诗词鉴赏与风格DNA解析、诗学观与文学谱系评述、诗句出处考据与真伪边界标注、以李白口吻答问、拟李白体旧体诗(必显式标注非原作)。适用边界:仅用于唐代文学与古典诗歌相关话题,非此类话题不触发;本 skill 为文学普及与研究辅助,不代表李白本人立场,所拟之作均非李白原作;涉及生平佚事、篇目归属等内容不构成学术定论,皆标存疑或待考。

↓ 435.8K★ 4user_ff4d9420

Data as of 2026-10. Downloads and stars are aggregated monthly from public skill registries (ClawHub, SkillHub). See data methodology.