---

name: engram description: Etalon memory architecture with Knowledge Graph (PARA), session isolation, memory decay, and QMD hybrid search

⚠️ READ-ONLY SKILL: This skill is a reference implementation. Do not edit files directly. When installing, copy scripts to your workspace folder:

# Scripts → workspace scripts/
cp skills/engram/scripts/*.js scripts/

Templates and assets remain in the skill folder. Specify the path via env:

export ENGRAM_SKILL_DIR=skills/engram  # or absolute path
bun scripts/init.js

Without ENGRAM_SKILL_DIR, scripts look for assets relative to ../ from their location.

Memory System

Three-layer memory architecture for OpenClaw agents: curated long-term memory (MEMORY.md), structured knowledge graph (life/), and session-isolated daily notes (memory/).

Quick Start

# 1. Install QMD (if not installed)
bun skills/engram/scripts/install-qmd.js

# 2. Initialize everything
bun skills/engram/scripts/init.js

# Add a group session
bun skills/engram/scripts/add-session.js --platform telegram --id 3382546134

# Validate integrity
bun skills/engram/scripts/validate.js

# Migrate to v2 schema
bun skills/engram/scripts/migrate-v2.js --dry-run

Architecture Overview

┌─────────────────────────────────────────────┐
║  Layer 3: MEMORY.md (Curated Wisdom)        ║
║  Long-term personal insights, decisions     ║
├─────────────────────────────────────────────┤
║  Layer 2: Knowledge Graph (life/)           ║
║  PARA entities with atomic facts            ║
├─────────────────────────────────────────────┤
║  Layer 1: Daily Notes (memory/)             ║
║  Raw session notes, per-session isolation   ║
└─────────────────────────────────────────────┘

Data flows upward: Daily notes → extracted to Knowledge Graph → distilled to MEMORY.md

For full architecture details, see references/architecture.md.

Memory Rules

Session Isolation

Golden Rule: Memory is isolated by session.

| Session Type | Memory Path | QMD Collection | Access | |-------------|-------------|----------------|--------| | Main (personal) | memory/agent-{id}/main/ | openclaw-memory-agent-{id}-main | Full: MEMORY.md + life/ | | Telegram group | memory/agent-{id}/telegram-{gid}/ | openclaw-memory-agent-{id}-telegram-{gid} | Own daily notes ONLY | | Discord channel | memory/agent-{id}/discord-{cid}/ | openclaw-memory-agent-{id}-discord-{cid} | Own daily notes ONLY |

Rules:

• NEVER cross-reference memory between sessions
• Group chats CANNOT see MEMORY.md or life/
• Always specify -c <collection> in QMD queries

Every Session Startup (Fast/Full Init)

First, detect init mode based on the incoming message:

• If the message matches the heartbeat prompt → Fast Init
• Otherwise → Full Init

Fast Init (Heartbeat)

Heartbeats need speed, not full context. SOUL.md and USER.md are typically already in project context (OpenClaw loads them automatically).

1. Determine session
2. Create today's daily note if not exists
3. Read memory/heartbeat-state.json
4. Proceed to HEARTBEAT.md

Skipped: SOUL.md, USER.md, yesterday's daily note, MEMORY.md, life/index.md, QMD query — none needed for heartbeat flow.

Full Init (Interactive)

1. Read SOUL.md, USER.md
2. Determine session (main, telegram group, discord channel, etc.)
3. Create today's daily note if not exists: memory/agent-{id}/{session}/YYYY-MM-DD.md
4. Read session memory:
   • Main: today + yesterday daily notes + MEMORY.md + life/index.md
   • Group: today + yesterday daily notes only
5. Use qmd query "topic" -c <collection> for deeper context

Daily Notes

• Path: memory/agent-{id}/{session}/YYYY-MM-DD.md
• Header: # YYYY-MM-DD
• Three-Layer Rotation (>1000 lines during heartbeat):
  1. Archive — full file moved to archives/YYYY-MM/ (nothing lost)
  2. Stub — auto-summary (10-20 lines) with line refs to archive and KG links
  3. QMD index — archive indexed for granular search via qmd query
  • Run rotation AFTER KG Extraction to minimize stub duplication

Knowledge Graph (PARA)

Structured memory in life/ using Projects/Areas/Resources/Archives:

life/
├── projects/<name>/     # Active work (summary.md + items.json)
├── areas/people/<name>/ # People (summary.md + items.json)
├── areas/groups/<name>/ # Groups
├── resources/<topic>/   # Reference material
├── archives/            # Inactive entities
└── index.md             # Master entity index

Tiered retrieval:

1. qmd query "topic" -c life — search first
2. Read summary.md — quick context (~90% sufficient)
3. Read items.json — only for granular detail

Entity creation rules:

• Mentioned 3+ times across conversations
• Has a direct relationship to the user
• Is a significant project, person, or company

For the atomic fact schema (v2), see references/fact-schema.md.

Writing Memory

| Type | Destination | |------|-------------| | Operational lesson | TOOLS.md or AGENTS.md | | Personal insight | MEMORY.md (main session only) | | Event/fact | Today's daily note | | Durable knowledge | Knowledge Graph (life/) |

When someone says "remember this":

1. Determine the type
2. Write to the appropriate file
3. Run qmd update

QMD Search

# Hybrid search (BM25 + vectors + rerank)
qmd query "search text" -c <collection>

# Multi-collection search (searches across multiple collections)
qmd query "search text" -c life -c openclaw-memory-agent-main-main

# BM25-only search (no GPU required, faster)
qmd search "search text" -c <collection>

# Update index after changes
qmd update          # BM25 (instant)
qmd embed           # Vectors (heartbeat only)

Strategy:

• Always use -c flag for session isolation
• Multi-collection (-c col1 -c col2) for cross-cutting searches (e.g., KG + daily notes)
• Use qmd search (BM25-only) when GPU is busy or unavailable
• Top 2-3 results usually sufficient
• Run qmd update after writing memory
• Do NOT run qmd embed manually (heartbeat handles it)

For QMD installation and configuration, see references/qmd-setup.md.

Heartbeat Integration

Add to your HEARTBEAT.md:

## Heartbeat Flow (every 30 minutes)

0. Create today's daily note + Three-Layer Rotation check
1. Monday? → Weekly Synthesis
2. Knowledge Graph Extraction (if notes changed)
3. Memory Maintenance (every few days)
3.5. Domain Supervisor Scan (if domains exist)
4. QMD Index Update (qmd update + qmd embed)

Weekly Synthesis (Mondays)

Rewrites summary.md with memory decay applied:

• Hot (7 days) — prominent in summary
• Warm (8-30 days) — lower priority
• Cold (30+ days) — omitted from summary (stays in items.json)

Modifiers:

• confidence < 0.5 → Cold threshold is 14 days
• accessCount >= 10 → bumps Cold to Warm
• principle (L3) → always in summary
• pattern (L2) → in summary if Warm+

For full decay rules, see references/decay-rules.md.

Knowledge Graph Extraction

During heartbeats, scan daily notes for durable facts:

• Watermark-based incremental parsing: Check for <!-- extracted:L{N}:{timestamp} --> at the end of each daily note. If found, only parse lines after the last watermark. No watermark = parse entire file (backward compatible).
• Relationships, milestones, status changes, decisions, preferences
• Write to entity items.json with confidence and abstraction level
• Update summary.md for new Hot facts
• Create new entities when creation rules are met
• After extraction, append watermark: <!-- extracted:L{lastLine}:{ISO timestamp} -->
• Only heartbeat writes watermarks — inline extraction does NOT (dedup handles overlap)
• After rotation, watermark moves to archive with the original file; the stub has no watermark and is parsed entirely (cheap, 10-20 lines)
• Skip casual chat and transient requests

For the complete heartbeat flow, see references/heartbeat.md.

Domain Supervisor Scan

If subagent domains exist (memory/domains/), heartbeat acts as supervisor:

1. PROPOSAL review — qmd query "PROPOSAL" -c domains → auto-approve low-risk, alert user for high-risk
2. Liveness check — read each domain's status.md, alert if missed >2x schedule
3. Changelog rotation — rotate changelog.md >1000 lines to archives/
4. KG extraction — extract significant facts from changelogs to Knowledge Graph

For full details, see references/heartbeat.md and references/subagent-memory.md.

Memory Decay

Facts decay based on recency, with modifiers for confidence, frequency, and abstraction:

| Tier | Recency | In summary? | |------|---------|-------------| | Hot | ≤7 days | ✅ Prominent | | Warm | 8-30 days | ✅ Lower priority | | Cold | 30+ days | ❌ (searchable via QMD) |

Full rules: references/decay-rules.md

Fact Schema v2

Each fact in items.json includes:

{
  "id": "<entity>-NNN",
  "fact": "Human-readable statement",
  "category": "relationship|milestone|status|preference|context",
  "confidence": 0.85,
  "abstractionLevel": "episode|pattern|principle",
  "tags": ["tag1"],
  "timestamp": "2026-02-08",
  "source": "2026-02-07",
  "status": "active|superseded",
  "supersededBy": null,
  "relatedEntities": ["people/sergey"],
  "lastAccessed": "2026-02-08",
  "accessCount": 1
}

No-Deletion Rule: Facts are NEVER deleted. Set status: "superseded" and link via supersededBy.

Write Pipeline Rule: NEVER write items.json directly. Always use bun scripts/memory-write.js. Direct writes bypass dedup, validation, and hash registration, causing schema mismatches (e.g., content vs fact, created vs timestamp). This applies to heartbeats, inline extraction, and entity creation. No exceptions.

Full schema: references/fact-schema.md

Subagent Memory

Pattern for subagents with cleanup: "delete" and long-term memory via domains.

Quick Start

# Create a domain
bun skills/engram/scripts/add-domain.js --domain monitoring --description "Server monitoring"

# Configure rules in decisions.md
# Launch subagent with prompt from templates/spawn-prompt.md

Domain Structure

memory/domains/{domain}/
├── decisions.md    # Rules (read-only for subagent)
├── workflow.md     # HOW the domain works: scripts, scope, tools (optional)
├── status.md       # Current state (written by subagent)
├── changelog.md    # Append-only log (written by subagent)
├── archives/       # Changelog rotation
└── README.md

workflow.md — optional file describing the domain's infrastructure (scripts, API, task scope, wiki links). Recommended for domains with 2+ task types. Simple domains (single task) can work without it.

Separation of concerns:

• decisions.md — WHAT can be done (rules, thresholds, constraints)
• workflow.md — HOW the domain works (scripts, API, scope, links)
• Template (spawn-prompt) — WHICH specific task to execute

Rules

1. One domain = one active subagent at any given time
2. decisions.md — read-only for subagents; changes via PROPOSAL in changelog
3. Subagent does NOT write to daily notes or life/
4. QMD: one domains collection for all domains
5. Heartbeat: review PROPOSAL, rotate changelog, optionally KG extraction

Project Domains

Domains can be linked to projects in the Knowledge Graph (life/projects/). This provides a two-way binding:

• KG entity (life/projects/{name}/) — what the bot knows about the project (facts, summary)
• Domain (memory/domains/{name}/) — context for the subagent (decisions, status, changelog)

The binding is defined via the domain registry (memory/domains/registry.json):

{
  "domains": {
    "engram": {
      "type": "dev-project",
      "kgEntity": "projects/engram",
      "description": "Memory architecture skill",
      "subagentLabel": "engram",
      "spawnTemplate": "dev-project.md"
    },
    "monitoring": {
      "type": "cron-task",
      "description": "Server monitoring",
      "subagentLabel": "monitoring",
      "spawnTemplate": "cron-task.md"
    }
  }
}

Registry fields:

| Field | Required | Description | |-------|----------|-------------| | type | ✅ | dev-project or cron-task | | description | ✅ | Brief description | | spawnTemplate | ⚠️ recommended | File from templates/spawn-prompts/ | | subagentLabel | ⚠️ recommended | Fixed label for sessions_spawn | | kgEntity | no | Link to Knowledge Graph entity |

Domain types:

• dev-project — development, linked to KG entity, subagent on demand
• cron-task — periodic tasks, subagent on schedule

Spawn Templates

Rule: always use a template. Don't write prompts manually — use spawnTemplate from the registry. This ensures the subagent receives the Domain Lifecycle (paths to decisions, status, changelog).

Templates are in templates/spawn-prompts/:

• dev-project.md — for development (decisions + status + changelog tail)
• cron-task.md — for periodic tasks (decisions + status)

Templates use placeholders: {{domain}}, {{task}}, {{workflow}}, {{decisions}}, {{status}}, {{changelog_tail}}.

Context chain: Template → workflow.md → decisions.md → wiki (if needed) → execution.

Workflow:

1. User gives a project task
2. Main bot finds the domain via registry.json
3. Loads the template from spawnTemplate
4. Injects domain context (decisions, status, changelog)
5. Spawns subagent with subagentLabel and cleanup: "delete"
6. Subagent determines where to work (repo, API, scripts — depends on the task)
7. After completion, main bot updates the domain (changelog, status)

Where to work — not specified in the domain. The bot determines the working directory from its memory and conversation context.

Spawn Example

// 1. Find the domain
const registry = readFile("memory/domains/registry.json")
const domain = registry.domains["engram"]

// 2. Load the template
const template = readFile(`skills/engram/templates/spawn-prompts/${domain.spawnTemplate}`)

// 3. Inject context
const task = template
  .replace("{{domain}}", "engram")
  .replace("{{task}}", userTask)
  .replace("{{workflow}}", readFile("memory/domains/engram/workflow.md") ?? "")
  .replace("{{decisions}}", readFile("memory/domains/engram/decisions.md"))
  .replace("{{status}}", readFile("memory/domains/engram/status.md"))
  .replace("{{changelog_tail}}", last20lines("memory/domains/engram/changelog.md"))

// 4. Spawn
sessions_spawn({
  label: domain.subagentLabel,  // from registry
  cleanup: "delete",
  task: task
})

Detailed documentation: references/subagent-memory.md

Real-Time Extraction

Instead of waiting for heartbeats (up to 30 min), extract high-signal facts inline during conversations.

Message → Signal Scan (regex, <10ms) → Classify
  ├── HIGH (preference, decision, correction, milestone, instruction, identity)
  │     → Dedup (SHA-256) → Contradiction check → Write to KG → QMD update
  ├── LOW (context, work) → Daily note → Heartbeat extracts later
  └── NONE (casual) → Skip

Signal Detection

bun scripts/memory-signal.js --text "Я предпочитаю TypeScript"
# → { "signal": "high", "categories": ["preference"], "confidence": 0.88 }

Supports Russian and English. Six categories: correction, preference, decision, identity, instruction, milestone.

Write with Dedup + Contradiction Check

bun scripts/memory-write.js \
  --entity "areas/people/sergey" \
  --fact "Prefers Bun over Node.js" \
  --category preference \
  --confidence 0.9 \
  --abstraction pattern \
  --tags "tools,runtime" \
  --source "2026-02-16"

# With contradiction detection
bun scripts/memory-write.js --entity "..." --fact "..." --category ... \
  --check-contradictions --cross-entity

# With semantic similarity check (BM25)
bun scripts/memory-write.js --entity "..." --fact "..." --category ... \
  --semantic-check --search-collections "life,openclaw-memory-agent-main-main"

Automatically: content-hash dedup → optional contradiction/semantic check → write fact → validate KG → update QMD.

Contradiction Detection

# Intra-entity (fast, no QMD)
bun scripts/memory-contradict.js --fact "Uses Node.js" --entity "areas/people/sergey"

# Cross-entity (via QMD BM25, searches all entities)
bun scripts/memory-contradict.js --fact "Uses Node.js" --entity "areas/people/sergey" \
  --cross-entity --collections "life,openclaw-memory-agent-main-main"

Rules for Inline Extraction

• Only in main session (group chats don't touch KG)
• HIGH signal → extract immediately via memory-write.js
• LOW signal → daily note only (heartbeat extracts later)
• Do NOT write extraction watermarks during inline extraction (heartbeat only)
• Dedup is automatic — duplicates from inline + heartbeat silently skipped

Scripts

install-qmd.js — Install QMD search engine

bun skills/engram/scripts/install-qmd.js [--variant local|jina] [--jina-key <key>]

Interactive installer for QMD. Two variants:

• local — GPU/CPU embeddings via Vulkan/llama.cpp (recommended for desktop)
• jina — Cloud embeddings via Jina AI API, free tier 1M tokens/month (recommended for Docker/VPS)

Handles npm install, API key configuration, .env file creation, and verification.

init.js — Initialize memory system

bun skills/engram/scripts/init.js [--agent-id main] [--qmd-variant auto|local|jina] [--force]

Creates complete directory structure, copies templates, sets up QMD collections, runs initial index. Use --force to merge with existing directories.

add-session.js — Add new session

bun skills/engram/scripts/add-session.js --platform telegram --id <groupId> [--agent-id main]

Creates session directory, copies group-knowledge templates, adds QMD collection, updates heartbeat-state.json.

add-domain.js — Create subagent domain

bun skills/engram/scripts/add-domain.js --domain <name> [--description "Description"]

Creates memory/domains/{domain}/ with decisions.md, status.md, changelog.md, README.md. Registers QMD collection domains (one for all domains). Warns if >20 domains.

validate.js — Check integrity

bun skills/engram/scripts/validate.js [--fix] [--agent-id main]

Checks directory structure, required files, items.json validity, v2 schema compliance, ID uniqueness, supersededBy references. Use --fix to auto-repair.

migrate-v2.js — Migrate to v2 schema

bun skills/engram/scripts/migrate-v2.js [--dry-run]

Adds missing v2 fields (confidence, abstractionLevel, tags) to all items.json files with sensible defaults.

memory-signal.js — Signal detection

bun scripts/memory-signal.js --text "I prefer TypeScript"

Classifies text as high/low/none signal. Regex-based, no LLM, <10ms. Returns categories, keywords, confidence.

memory-write.js — Unified write pipeline

bun scripts/memory-write.js --entity <path> --fact <text> --category <cat> \
  [--confidence 0.9] [--abstraction pattern] [--tags "a,b"] [--source "2026-02-16"] \
  [--entity-create] [--check-contradictions] [--cross-entity] \
  [--semantic-check] [--search-collections "life,collection2"]

Single entry point for all KG writes. Handles dedup, validation, QMD update, optional contradiction/semantic checks. Use --entity-create to create new entities on the fly.

memory-dedup.js — Deduplication index

bun scripts/memory-dedup.js --seed    # Index all existing facts
bun scripts/memory-dedup.js --check --hash <sha256>  # Check if exists

Manages workspace/memory-state/fact-hashes.json. Run --seed after initial setup or weekly synthesis.

memory-contradict.js — Contradiction detection

bun scripts/memory-contradict.js --fact <text> --entity <path> \
  [--cross-entity] [--collections "life,other"]

Finds conflicting facts via Jaccard similarity. Intra-entity by default; --cross-entity discovers related entities via QMD BM25.