# Memory Orchestrator

> Three-layer memory system for OpenClaw agents with automatic Dream consolidation. Orchestrates LCM (lossless context), task files, and daily logs into a coherent architecture with periodic self-maintenance. Use when setting up a new OpenClaw instance's memory system, when the user asks about memory/context management, or when configuring plugins for persistent agent memory. Covers plugin installation, config, cron jobs, Dream consolidation, file conventions, and cross-platform (WhatsApp/webchat) context continuity.

- Skill: `enderfga/memory-orchestrator` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add enderfga/memory-orchestrator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/enderfga/memory-orchestrator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Enderfga (https://skillmd.com/u/enderfga)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/enderfga/memory-orchestrator

---


# Memory Orchestrator

Three-layer memory system that gives OpenClaw agents persistent, cross-session, cross-platform memory — with automatic Dream consolidation to prevent memory bloat.

## Architecture Overview

```
Layer 1: LCM (session-internal)    → auto DAG summaries, never loses context
Layer 2: Task files + daily logs (cross-session) → dense snapshots + event journal
Layer 3: MEMORY.md (long-term)     → curated high-level overview
```

Each layer serves a different density/scope tradeoff. See `references/architecture.md` for the full design.

## Quick Setup

### 1. Install plugins

```bash
# Lossless Context Management (session-internal, DAG summaries)
openclaw plugins install @martian-engineering/lossless-claw
```

### 2. Configure

Apply this config patch (adjust `summaryModel` to your setup):

```json
{
  "plugins": {
    "slots": {
      "contextEngine": "lossless-claw"
    },
    "entries": {
      "lossless-claw": {
        "enabled": true,
        "config": {
          "freshTailCount": 32,
          "contextThreshold": 0.75,
          "incrementalMaxDepth": -1,
          "ignoreSessionPatterns": ["agent:*:cron:**"],
          "summaryModel": "anthropic/claude-haiku-4-5"
        }
      },
      "memory-core": {
        "enabled": true,
        "config": {
          "dreaming": { "enabled": true, "frequency": "0 4 * * *" }
        }
      }
    }
  },
  "agents": {
    "defaults": {
      "compaction": {
        "mode": "safeguard",
        "memoryFlush": { "enabled": true }
      }
    }
  }
}
```

### 3. File structure

Create these directories in your workspace:

```
workspace/
├── MEMORY.md              # Long-term curated memory (weekly updates)
├── memory/
│   ├── YYYY-MM-DD.md      # Daily logs
│   └── tasks/             # Dense task snapshots
│       ├── project-a.md
│       └── project-b.md
```

### 4. Cron jobs (recommended)

Set up these automated memory maintenance jobs:

- **Daily Memory Sync** (23:00) — Scan all sessions, write daily log, backfill missing task files. Use a cheap/fast model (e.g., Gemini Flash).
- **Weekly Compound** (Sunday 22:00) — Distill weekly logs into MEMORY.md updates.

See `references/cron-templates.md` for ready-to-use cron job definitions.

## Layer Details

### Layer 1: LCM (Lossless Context Management)

**What**: Every message persisted to SQLite. Old messages auto-summarized into a DAG. Context assembled each turn = summaries + recent raw messages.

**When it kicks in**: Context usage hits `contextThreshold` (default 75%).

**Agent tools**: `lcm_grep` (search history), `lcm_describe` (inspect summary), `lcm_expand` (drill into details).

**Key config**: See `references/lcm-tuning.md` for parameter explanations.

### Layer 2: Task Files

**What**: Dense markdown snapshots of active work, stored in `memory/tasks/`.

**Convention**:
- Create when user says "write task" or at end of substantial work sessions
- Naming: kebab-case (`ta-leaderboard.md`, `openvla-exp.md`)
- Load on-demand only (when task is mentioned)
- Mark `[DONE]` when complete, never delete
- Daily Sync auto-creates lightweight tasks for uncovered work

### Layer 2: Daily Logs

**What**: `memory/YYYY-MM-DD.md` — chronological event journal.

**Convention**:
- Load today + yesterday at session start
- Append throughout the day
- Source for weekly compound → MEMORY.md

### Layer 3: MEMORY.md

**What**: Curated high-level overview — project status, lessons learned, weekly summaries.

**Convention**:
- Update weekly (manually or via Weekly Compound cron)
- Only load in main/private sessions (contains personal context)
- Periodically review and prune stale entries

## Cross-Platform Continuity

For setups with multiple chat surfaces (e.g., WhatsApp + webchat):

1. **Webchat idle reset**: Set `session.resetByChannel.webchat.idleMinutes` to 7+ days to prevent orphaned sessions
2. **Task handoff**: Complete deep work on webchat → "write task" → continue on WhatsApp seamlessly
3. **Daily Sync backfill**: Catches webchat work that wasn't manually task-filed

See `references/cross-platform.md` for the full WhatsApp + webchat cooperation model.

## Dream Consolidation

### OpenClaw Native Dreaming (2026.4.5+)

OpenClaw now ships with built-in dreaming under `memory-core`. Enable it with:

```json
{
  "plugins": {
    "entries": {
      "memory-core": {
        "enabled": true,
        "config": {
          "dreaming": { "enabled": true, "frequency": "0 4 * * *" }
        }
      }
    }
  }
}
```

Native dreaming handles: recall promotion (weighted short-term → long-term), decay aging, daily-note chunking, dreams.md output, and `/dreaming` command.

### Custom Dream Cron — Complements Native

For MEMORY.md line-count enforcement and task file archival, the custom Dream cron is still valuable. The four-step cycle runs weekly on-demand:

1. **Orientation** — Read MEMORY.md, assess current state
2. **Consolidation** — Update statuses, remove completed items, compress old reports, enforce absolute dates
3. **Task File Review** — Archive completed tasks to `memory/tasks/archive/`
4. **Index** — Ensure MEMORY.md stays under target line count, update timestamp

See `references/dream-consolidation.md` for details and `references/cron-templates.md` for the cron job definition.

### Absolute Date Rule

All memory files (MEMORY.md, task files) must use `YYYY-MM-DD` format. Never write "today", "yesterday", or "last week" — these become meaningless after the session ends.

## Maintenance

The Dream cron handles most maintenance automatically. For manual maintenance during heartbeats:
1. Review recent `memory/YYYY-MM-DD.md` files
2. Distill significant events into MEMORY.md
3. Check `memory/tasks/` for stale ACTIVE tasks (>3 days no update)
4. Prune outdated MEMORY.md entries

## Troubleshooting

- **LCM not triggering**: Check if context window is large enough (200k+ may rarely hit 75%). Lower `contextThreshold` to 0.6 for testing.
- **Task files empty**: Daily Sync may not be running. Check cron job status.
- **Webchat context lost**: Verify `resetByChannel.webchat.idleMinutes` is set high enough.