MemoryRouter ⚡#

Never lose context. Never forget decisions. Never repeat mistakes.

The memory problem is the single biggest reason agents feel dumb over time. Not the models. Not the prompts. Memory management.

MemoryRouter fixes it with one tool, zero dependencies, zero setup.

The Problem — In 30 Seconds#

Agent memory files grow unbounded. After a week, you've got thousands of lines across dozens of files. Every session loads everything — even when the user asks "what's the weather?"

That's tokens burned on irrelevant memories, every interaction, forever.

Then compaction kicks in and details vanish.

The bottleneck isn't the model. It's what the model gets to see.

Why Memory Fails#

Failure ModeCauseFixForgets everythingLoads irrelevant files, context window fills upSmart manifest — load only what mattersRepeats mistakesLessons not captured or loadedEntity index + audit for conflictsRepeats workNo session state persistenceWAL protocol — write state before respondingSlow responsesLoads 50+ files when only 3 matterToken budgeting — cap context windowDuplicates everywhereNo automated cleanup--audit finds high-similarity pairs

The Architecture#

text

┌──────────────────────────────────────────────────────┐
│              MEMORYROUTER ⚡                          │
├──────────────────────────────────────────────────────┤
│                                                      │
│  ┌──────────────┐    ┌──────────────┐               │
│  │  AUTO-TIER   │    │  MANIFEST    │               │
│  │  MEMORY.md   │ →  │  GENERATOR   │               │
│  │              │    │              │               │
│  │ Core (always │    │ Required:    │               │
│  │ loaded)      │    │ MEMORY.md    │               │
│  │ + Archive    │    │ Recent logs  │               │
│  │ (on demand)  │    │ Boosted:     │               │
│  └──────────────┘    │ entity files │               │
│                      └──────┬───────┘               │
│                             │                        │
│              ┌──────────────▼──────────────┐         │
│              │  ENTITY RESOLUTION          │         │
│              │  "alice" → preferences.md,         │         │
│              │  notes.md, decisions.md│         │
│              └──────────────┬──────────────┘         │
│                             │                        │
│              ┌──────────────▼──────────────┐         │
│              │  TOKEN BUDGET FILTER        │         │
│              │  20K budget → 7 files       │         │
│              │  100K budget → 16 files     │         │
│              └──────────────┬──────────────┘         │
│                             │                        │
│              ┌──────────────▼──────────────┐         │
│              │  AGENT LOADS ONLY WHAT      │         │
│              │  MATTERS → 70-85% REDUCTION │         │
│              └─────────────────────────────┘         │
└──────────────────────────────────────────────────────┘

⚠️ Important Warnings#

Automatic File Modification#

MemoryRouter automatically modifies MEMORY.md and archive files under memory/ during --tier, --audit, and --restore operations. These changes are not reversible unless you have a backup. Always run --dry-run first to preview changes.

Destructive Flags Requiring Confirmation#

• --tier rewrites MEMORY.md permanently
• --restore overwrites MEMORY.md with backup content
• --cleanup with auto-remove can delete archived entries
• These operations require explicit --confirm / --force flags and will refuse to run without them

Filesystem Access#

MemoryRouter reads and writes memory/ files and directories. Ensure proper filesystem permissions and avoid running in untrusted shared directories.

Quick Start#

Fix a bloated MEMORY.md (requires confirmation)#

bash

node skills/memory-router/memory-router.js --tier --confirm

Output:

text

[memory-router] MEMORY.md: 7809 lines, 309254 chars
[memory-router] ✅ Pre-tier backup: memory/backups/MEMORY-backup-2026-05-23-1234567890.md (7809 lines saved)
[memory-router] Tiered: 202 core sections, 1002 archived
[memory-router] MEMORY.md reduced from 7809 lines to 2222 lines

Preview tiering without making changes#

bash

node skills/memory-router/memory-router.js --tier --dry-run

Output:

text

[memory-router] MEMORY.md: 7809 lines, 309254 chars
[memory-router] ── DRY RUN ──
[memory-router] Would archive: memory/active/MEMORY-archive-2026-05-23.md
[memory-router] Would reduce MEMORY.md from 7809 lines to ~2222 lines
[memory-router] Core sections: 202, Archive sections: 1002
[memory-router] ✅ No files were modified.

Restore MEMORY.md from latest backup#

bash

node skills/memory-router/memory-router.js --restore --force

⚠️ Destructive — overwrites MEMORY.md. Requires --force flag. Extracts only the original content from the backup (metadata headers are stripped). Always creates a backup of the current MEMORY.md before overwriting.

Generate a smart manifest in one command#

bash

node skills/memory-router/memory-router.js --compact

Creates memory/memory-manifest.json — the agent's shopping list of what to load.

With entity boosting#

bash

node skills/memory-router/memory-router.js --compact --query "alice"

Files linked to "alice" get priority. No AI, no embeddings — just fast entity resolution.

With a token budget#

bash

node skills/memory-router/memory-router.js --compact --budget 20000

Only loads files that fit within 20K tokens. Tight budgets load core only. Generous budgets load archive on demand.

Other commands#

bash

node skills/memory-router/memory-router.js --audit      # Find duplicates & conflicts
node skills/memory-router/memory-router.js --status      # Health overview
node skills/memory-router/memory-router.js --entity add alice person preferences.md

How It Works#

The Core Insight#

Memory management is the bottleneck, not model capability. Every agent system hits the same wall — context window fills up, everything loads, irrelevant memories dilute the signal.

MemoryRouter takes a routing approach rather than a compression approach:

1. Auto-tiering splits MEMORY.md into core sections (identity, preferences, boundaries — always loaded) and archive sections (everything else, loaded on demand)
2. Manifest generation creates a per-session file list: which files to load, which to skip, which to boost
3. Entity-aware boosting links people, projects, and systems to files — search for "alice" → load preferences.md first
4. Token budgeting caps how many files load based on available context window

The Flow#

text

User query → --compact --query "alice"
              ↓
         Manifest generated
              ↓
         Load required files (MEMORY.md, recent daily logs)
              ↓
         Boost entity-matched files (preferences.md, notes.md)
              ↓
         Agent loads only what matters → 70-85% context reduction

The 4 Engines#

⚡ Engine 1: Auto-Tier ( --tier)#

When MEMORY.md exceeds configurable thresholds (default: 500 lines or 25KB), splits into:

• Core file — Identity, preferences, relationships, projects, patterns, boundaries
• Archive files — Everything else, stored with timestamps in memory/active/

Safety features (v2):

• Pre-tier backup — Every --tier creates a full backup in memory/backups/ before modifying MEMORY.md
• --dry-run — Preview what tiering would do without making changes
• --confirm — Required for destructive writes (no auto-tier without explicit confirmation)
• Minimum core size — Aborts if core sections would fall below minCoreLines/ minCoreChars thresholds
• --restore --force — Restore MEMORY.md from the latest backup (requires --force flag)

Sections are classified by header keywords (identity, preferences, etc.) or by content heuristics.

📋 Engine 2: Manifest Generator ( --compact)#

Creates memory/memory-manifest.json with a file list:

json

{
  "generated": "2026-05-21",
  "files": [
    { "path": "MEMORY.md", "tier": "core", "required": true, "size": 50346 },
    { "path": "memory/2026-05-21.md", "tier": "recent", "required": true, "size": 2034 },
    { "path": "self-improving/memory.md", "tier": "domain", "required": false, "size": 670 }
  ]
}

Options:

FlagDescription--query "text"Entity-aware boosting — files linked to matching entities get priority--budget NToken budget — only loads files that fit within N tokens

🔍 Engine 3: Audit Scanner ( --audit)#

Scans all memory files for:

• High-similarity pairs — Files with >70% text overlap (potential duplicates)
• Revision keywords — "revised", "updated", "changed", "no longer", "actually", "correction" (facts that may have been superseded)

Output: memory/memory-audit-report.md

🏥 Engine 4: Health Monitor ( --status)#

Quick snapshot of MEMORY.md line/char counts, file counts per directory, archive count, manifest status, entity index size, and WAL state.

Additional Tools#

Entity Index ( --entity)#

CommandDescription--entity add <name> <type> <files...>Add entity (e.g., alice person preferences.md)--entity listList all entities--entity search <query>Search entities (direct and fuzzy match)

WAL Protocol ( --wal)#

CommandDescription--wal initInitialize SESSION-STATE.md with template--wal getShow current session state--wal update <section> --content "<text>"Update a section

Configuration#

Edit config.json to customize behavior:

json

{
  "thresholds": {
    "memoryMdMaxLines": 500,
    "memoryMdMaxChars": 25000,
    "tierArchiveMinAgeDays": 3,
    "auditMaxFiles": 50
  },
  "tiering": {
    "archiveDir": "memory/active",
    "retentionDays": 90,
    "keepInMemory": [
      "identity", "preferences", "relationships",
      "projects", "patterns", "boundaries", "key_facts"
    ]
  },
  "manifest": {
    "generateOnTier": true,
    "manifestPath": "memory/memory-manifest.json"
  },
  "audit": {
    "reportPath": "memory/memory-audit-report.md",
    "duplicateThreshold": 0.7,
    "conflictKeywords": [
      "revised", "updated", "changed", "no longer",
      "actually", "correction", "mistake"
    ]
  }
}

SettingDefaultDescriptionmemoryMdMaxLines500Auto-tier trigger (lines)memoryMdMaxChars25000Auto-tier trigger (characters)tierArchiveMinAgeDays3Minimum age before archivingretentionDays90⚠️ Archive retention period. Files older than this are candidates for deletion (irreversible). Set to 365+ until confident.keepInMemorysee aboveHeaders/keywords that stay in coregenerateOnTiertrueAuto-generate manifest after tieringduplicateThreshold0.7Similarity score to flag as duplicateconflictKeywordssee aboveWords that signal fact revisionbackupDir``memory/backupsDirectory for pre-tier backupsminCoreLines20Abort if core sections below thisminCoreChars1000Abort if core chars below this

Agent Memory Loading Protocol#

When the agent wakes up, use the manifest instead of loading all memory files:

1. Read memory/memory-manifest.json
2. Load all required: true files
3. For required: false files, use memory_search to check relevance
4. Load only the top 3-5 most relevant optional files

Result: 70–85% context reduction — load what matters, skip the rest.

Heartbeat Integration#

Add to your HEARTBEAT.md:

markdown

### ⚡ MemoryRouter (SAFE commands only)

- Run `node skills/memory-router/memory-router.js --compact` to update manifest ✅ safe
- Run `node skills/memory-router/memory-router.js --audit` to check for issues ✅ safe
- Run `node skills/memory-router/memory-router.js --status` for health overview ✅ safe
- ⚠️ Do NOT auto-run `--tier` or `--restore` during heartbeats
- `--tier --dry-run` is side-effect free (no files created)
- For tiering: run `--tier --dry-run` manually, review output, then `--tier --confirm`

Performance#

MetricResultTiering speed (8K lines)25msTiering speed (15K lines)30msToken reduction96% (7,809 → 222 lines)File count reduction53 → 15 filesMemory footprint~2MB (Node.js runtime)

Comparison#

ApproachTokens SavedSetup EffortMaintenancePrivacyRaw file injection0%NoneManual✅**MemoryRouter70–85%NoneAutomated✅**Obsidian vault40–60%HighMedium⚠️ CloudVector DB (ChromaDB)70–85%Very HighHigh✅mem070–85%HighMedium⚠️ Cloud

MemoryRouter gives you the best token savings of the vector DB approach with zero setup effort.

Entity Naming#

Pick one canonical name per entity and reuse it consistently:

• Use full descriptive names: "machine learning" not "ML", "JavaScript" not "JS"
• Same string after lowercasing = same entity. Different strings = different entities
• Call --entity search periodically to verify your index

Examples:

bash

# ✅ Good — consistent, descriptive
node memory-router.js --entity add alice person preferences.md
node memory-router.js --entity add openclaw system AGENTS.md

# ❌ Bad — inconsistent, ambiguous
node memory-router.js --entity add alice person preferences.md
node memory-router.js --entity add Alice person notes.md
node memory-router.js --entity add JS person docs.md

Manifest Format#

The manifest JSON tells the agent which files to load:

json

{
  "generated": "2026-05-21",
  "version": 2,
  "query": "memory management",
  "budget": 20000,
  "files": [
    {
      "path": "MEMORY.md",
      "tier": "core",
      "required": true,
      "size": 50346
    },
    {
      "path": "memory/2026-05-21.md",
      "tier": "recent",
      "ageDays": 0,
      "required": true,
      "size": 2034
    },
    {
      "path": "self-improving/memory.md",
      "tier": "domain",
      "required": false,
      "size": 670
    }
  ]
}

FieldTypeDescriptiontierstringcore, recent, domain, or archive``requiredboolAlways load this filesizeintFile size in bytesboostedboolEntity match — higher priorityentityMatchstringEntity name that matchedloadboolIncluded under budget modebudgetUsedintTotal tokens loaded (budget mode)budgetEfficiencystringPercentage remaining (budget mode)

Troubleshooting#

ProblemFix"No MEMORY.md found"Create one: echo "# MEMORY.md" > MEMORY.md"Memory directory not found"Create it: mkdir -p memoryTiering not workingCheck thresholds — if under 500 lines and 25KB, nothing happens (by design)Manifest shows wrong filesRun --compact again — it regenerates fresh each timeEntity search returns nothingAdd entities first: --entity add <name> <type> <files...>Budget too smallCore files always load. Budget only controls optional files.

⚠️ SAFETY — READ BEFORE USING#

Destructive Operations#

--tier and --restore permanently modify user memory files.

• --tier rewrites MEMORY.md, moving sections to archive files
• --restore overwrites MEMORY.md with backup content
• Both require explicit confirmation ( --confirm / --force) — they will refuse to run without it
• Always run --tier --dry-run first to preview what will change
• Pre-tier backups are created only when actually tiering (not in dry-run mode)

Archive Retention — Data Loss Risk#

The retentionDays config (default: 90) marks archived files as candidates for deletion after that period. This is irreversible — once deleted, archived memory sections cannot be recovered.

Before enabling retention:

1. Back up your entire memory/ directory
2. Set retentionDays to a large value (365+) until you're confident
3. Monitor what gets deleted before reducing the value

Safe vs Unsafe Commands#

CommandSafe in automation?Modifies files?--compact✅ YesWrites memory-manifest.json (generated output)--audit✅ YesWrites memory-audit-report.md (generated report)--status✅ YesNo file writes--entity add⚠️ Yes — but writes entity indexYes (persists entity → file mapping)--entity list/search✅ YesNo file writes--wal init⚠️ Yes — but creates SESSION-STATE.mdYes (creates new file)--wal get✅ YesNo file writes--wal update⚠️ Yes — but modifies SESSION-STATE.mdYes (updates session state)--tier --dry-run✅ YesNo file writes (side-effect free)--tier --confirm❌ No — rewrites MEMORY.md + creates backupYes--restore --force❌ No — overwrites MEMORY.mdYes

--compact, --audit, --status, --entity list/search, --wal get, and --tier --dry-run are safe for unattended/heartbeat use.

--entity add, --wal init, and --wal update write persistent files — review before automating.

--tier --confirm and --restore --force are destructive — never automate.

Security#

Built with defense-in-depth:

• Path validation — rejects file paths outside workspace root
• Regex escaping — prevents injection in WAL section names
• Symlink protection — refuses to read/write symlinks
• Size limits — 10MB max file size
• Entity name validation — alphanumeric + hyphens/underscores only
• Content sanitization — prevents header injection in WAL updates
• Audit keyword validation — rejects regex metacharacters in conflict keywords

Design Principles#

1. Safe by default — Destructive operations require explicit flags ( --confirm, --force)
2. No external dependencies — Pure Node.js, no npm packages
3. Configurable — Thresholds, keywords, retention policies
4. Transparent — Generates reports, nothing is silently deleted
5. Reversible — Archives use dated filenames, original content preserved
6. Privacy-first — All local, no cloud APIs