Project Trident: Four-Tier Persistent Memory Architecture

The Problem

AI agents forget. Every session starts blank. Important context—corrections, decisions, patterns, relationships—evaporates. Most agent frameworks treat memory as an afterthought: flat files or vector databases without intelligent curation. The result: agents that repeat mistakes, lose critical context, and can't develop genuine identity over time.

The Solution: Four Layers of Durability

Trident models memory like computer storage—with an additional intelligence layer:

The result: an agent that never forgets what matters, develops genuine personality, maintains continuity through crashes, and recovers from catastrophic failure.

What Each Layer Does

New to Trident? → Read references/trident-lite.md
No Docker required. Works on Windows, Mac, Linux, and any VPS out of the box.

Already have MEMORY.md, SOUL.md, or custom memory files? → Run the migration script.

Want Docker-based semantic search? → Read references/deployment-guide.md

Deploying to VPS? → See Layer 2 (Backup & Updates) section.

Cost-conscious? → Read references/cost-calculator.md

The Four-Tier Architecture

┌─────────────────────────────────────────────────────────────┐
│ Conversation Input (user messages, tool results, events)    │
└──────────┬──────────────────────────────────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────┐
│ LAYER 0: LCM (Lossless Context Management) — RAM            │
│ ├─ SQLite persistence (every message)                       │
│ ├─ DAG lineage tracking                                     │
│ └─ Never loses a message, even after compaction             │
└──────────┬──────────────────────────────────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────┐
│ LAYER 0.5: Signal Router (Cron Agent) — SSD                 │
│ ├─ Every 10–30 min: read daily logs (WAL protocol)         │
│ ├─ Classify signals: corrections, decisions, facts, self   │
│ ├─ Route to Layer 1 buckets                                │
│ ├─ Integrated with HEARTBEAT (job checks, trading, etc.)   │
│ └─ Cost: $0–$1.44/day depending on model & interval        │
└──────────┬──────────────────────────────────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────┐
│ LAYER 1: Hierarchical Memory Buckets — HDD                  │
│ ├─ MEMORY.md (curated long-term)                           │
│ ├─ memory/daily/ (raw episodic logs)                       │
│ ├─ memory/self/ (personality, beliefs, identity)           │
│ ├─ memory/lessons/ (learnings, mistakes, tool quirks)      │
│ ├─ memory/projects/ (active work, recurring tasks)         │
│ ├─ memory/semantic/ (knowledge, facts, technical)          │
│ └─ .md format (human-readable, Git-compatible)             │
│                                                             │
│ Optional: memory/heartbeat/ for domain-specific tracking   │
│ ├─ job-search.md (career info, hot leads, follow-ups)     │
│ ├─ trading/portfolio.md (positions, P/L, strategy)        │
│ ├─ subagent-status.md (running agents, models, tasks)     │
│ └─ time-sensitive.md (reminders, deadlines, TTL items)    │
└──────────┬──────────────────────────────────────────────────┘
           │ (optional upgrade)
           ▼
┌─────────────────────────────────────────────────────────────┐
│ LAYER 1.5: Semantic Recall (Optional) — Add at 50K msgs    │
│ ├─ Qdrant vector search (Docker, binary, or cloud)         │
│ ├─ FalkorDB entity graphs (Docker, Redis, or cloud)        │
│ └─ Pre-turn context injection via Layer 0.5                │
└──────────┬──────────────────────────────────────────────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────┐
│ LAYER 2: Backup & Updates Subsystem — COLD                 │
│ ├─ Daily Git commits (memory version control)              │
│ ├─ VPS snapshots (point-in-time recovery)                 │
│ ├─ Pre-update snapshots (before OpenClaw upgrades)        │
│ ├─ Post-update healthchecks (verify all systems)          │
│ └─ Automated rollback on failure                           │
│                                                             │
│ Crons:                                                       │
│ ├─ 4:20 AM MDT: Layer 2 Unified Backup (Git + snapshot)   │
│ ├─ 2:00 AM MDT (Sun): Pre-update snapshot                 │
│ ├─ After update: Post-update healthcheck (21 tests)       │
│ └─ Cost: $0 (included in VPS + free Git hosting)          │
└─────────────────────────────────────────────────────────────┘

Platform Support

See references/platform-guide.md for platform-specific commands.

Core Layers (Detailed)

Layer 0: Lossless Context Management (LCM)

What it does:

• Captures every session message in SQLite
• Builds DAG (directed acyclic graph) of message lineage
• Preserves full history even after compaction

Why it matters:

• Baseline durability — recovery foundation
• Enables Layer 0.5 to reconstruct what happened
• Prerequisite: Enable lossless-claw plugin in openclaw.json

Cost: $0 (SQLite is built-in)

Key property: Summaries created by compaction link back to source messages via DAG — nothing is ever truly lost.

Layer 0.5: Signal Router & Heartbeat Integration

What it does:

• Runs every 10–30 minutes (configurable)
• Reads daily logs (WAL protocol) to detect signals
• Classifies signals into buckets (corrections, decisions, facts, self-signals)
• Routes to Layer 1 (.md files)
• Integrates HEARTBEAT checks:
  • Job search status (active leads, follow-ups, TTL items)
  • Trading positions (if trading hours: entry/exit/rebalance)
  • Subagent lifecycle (kill idle/completed, optimize models)
  • Time-sensitive reminders (e.g., "Outlier AI shutdown in X days")

Signal types and routing:

Model: Claude Haiku recommended (cost-optimized, ~95% accuracy for routing)

Cost: $0.72–$1.44/day depending on interval (15–30 min)

Security: SHA256 integrity verification prevents prompt injection

Template: scripts/layer0-agent-prompt-template.md (customizable)

Key features:

• Attention management — corrections flagged first (never skipped)
• Context awareness — reads SESSION-STATE.md for current priorities
• Deadline tracking — alerts for time-sensitive items (e.g., "X days until Outlier AI shutdown")
• Subagent optimization — suggests killing idle agents, downgrading Sonnet→Haiku
• Domain-specific routing — job search, trading, research, ops — each gets own bucket

Layer 1: Hierarchical Memory Buckets

What it does:

• Persistent .md file organization
• Human-readable, Git-compatible, debuggable
• Prevents noise accumulation (agentic curation vs auto-capture)

Core structure:

memory/
├── MEMORY.md              ← Curated long-term memory (high-signal, compressed)
├── daily/                 ← Raw episodic logs
│   ├── 2026-04-19.md
│   └── ...
├── self/                  ← Personality, identity, growth
│   ├── identity.md
│   ├── beliefs.md
│   └── growth.md
├── lessons/               ← Learnings, mistakes, tool quirks
│   ├── tools.md
│   ├── mistakes.md
│   └── workflows.md
├── projects/              ← Active work (one file per project)
│   ├── job-search.md
│   ├── infrastructure.md
│   └── ...
├── semantic/              ← Knowledge, facts, domain-specific
│   ├── biology.md
│   ├── crypto.md
│   └── ...
├── heartbeat/             ← Time-sensitive, domain-specific tracking
│   ├── job-search.md      ← Hot leads, follow-ups, deadline tracking
│   ├── trading/
│   │   ├── portfolio.md   ← Current positions, P/L, strategy
│   │   └── trades.md      ← Entry/exit log, reasoning
│   ├── subagent-status.md ← Running agents, models, TTL
│   └── time-sensitive.md  ← Reminders, deadlines, TTL items
└── layer0/                ← Signal router internals (security)
    ├── AGENT-PROMPT.md    ← Layer 0.5 instructions
    └── audit-log.md       ← Integrity verification log

Quality rule: Compress over accumulate; signal density over volume.

Cost: $0 (local .md files)

Layer 1.5: Semantic Recall (Optional)

When to add: Memory exceeds ~50K messages, or you need to query context from months ago.

Qdrant (Vector Search):

• Deployment: Docker, standalone binary, or Qdrant Cloud (free tier)
• Purpose: "What did we discuss about X six months ago?"
• Integration: Pre-turn context injection via Layer 0.5

FalkorDB (Entity Graphs):

• Deployment: Docker, Redis module, or FalkorDB Cloud
• Purpose: Entity relationship tracking ("Person X works at Company Y, mentioned on date Z")
• Integration: Graphiti MCP for automated extraction

These are optional. Trident core (Layers 0, 0.5, 1, 2) works perfectly without semantic recall.

Layer 2: Backup & Updates Subsystem

What it does:

• Daily Git commits — version control for memory
• VPS snapshots — point-in-time recovery (if deployed)
• Update resilience — pre-update snapshots + post-update healthchecks
• Automated rollback — revert to last known-good state on failure

Components:

Example backup cron (Layer 2 Unified):

# Cron: 4:20 AM MDT daily
# Task: Unified backup combining multiple backup operations

Steps:
1. git add -A
2. git commit -m "Daily snapshot: $(date)"
3. git push origin main
4. (Optional) curl Hostinger VPS API → create snapshot
5. Log results to memory/daily/$(date).md

Result: Memory version-controlled + point-in-time VM snapshot

Recovery procedures:

1. Memory corrupted → restore from Git:

   git log --oneline     # Find last good commit
   git reset --hard <commit>  # Restore
   

2. Deployment broken → restore VPS snapshot:

   # Via Hostinger dashboard or API
   # Rollback to pre-update state
   # Container restarts with clean image
   

3. Update failed → check post-update healthcheck log:

   cat memory/layer2/post-update-healthcheck.log
   # Identifies which subsystem failed
   # Suggests rollback if needed
   

Cost: $0 (Git free tier + VPS snapshots included)

Integration: Layer 0.5 + HEARTBEAT

Layer 0.5 is the operational manifestation of HEARTBEAT. Rather than manual heartbeat checks, Layer 0.5 runs automatically every 15 minutes and:

1. Reads HEARTBEAT.md for standing priorities
2. Checks career transition status — flags blocking items
3. Monitors trading hours — enables entry/exit at 6am–4pm MDT
4. Scans subagent lifecycle — kills idle agents, optimizes models
5. Tracks time-sensitive items — reminds before TTL expires
6. Routes all discoveries to Layer 1 — memory never loses these insights

Example signals detected by integrated Layer 0.5:

Daily 6:00 AM:
  ✅ "Check job email from Indeed" → memory/heartbeat/job-search.md
  ✅ "Outlier AI shutdown in 33 days" → memory/heartbeat/time-sensitive.md (urgent)

Daily 6:30 AM (trading hours):
  ✅ "BTC up 2% since yesterday" → memory/heartbeat/trading/portfolio.md
  ✅ Check stop/take-profit levels, liquidation risk

Daily 4:20 AM (during backup):
  ✅ Layer 2 runs git backup
  ✅ Captures all daily signals + updates

Every 15 min (Layer 0.5):
  ✅ Read daily logs
  ✅ Detect: corrections, decisions, self-signals
  ✅ Route to appropriate Layer 1 buckets
  ✅ Flag HEARTBEAT priorities
  ✅ Kill idle subagents (>30 min inactivity)
  ✅ Log actions to audit-log.md

Migration from Existing Memory

If you already have MEMORY.md, SOUL.md, or custom memory files:

chmod +x scripts/migrate-existing-memory.sh

# Dry run (preview changes, no files modified)
./scripts/migrate-existing-memory.sh --dry-run

# Live migration (automatic backup)
./scripts/migrate-existing-memory.sh

What it does:

1. Creates backup of all existing files → memory/migration-backup/
2. Creates Trident directory structure
3. Guides you through file routing (interactive)
4. Installs AGENT-PROMPT.md
5. Generates migration report

Safety guarantees:

• Dry-run mode previews all changes
• Originals always backed up (never deleted)
• You approve each routing decision

Security

Template Integrity

Layer 0.5 reads and executes memory/layer0/AGENT-PROMPT.md. A compromised prompt = compromised routing.

Setup:

chmod +x scripts/template-integrity-check.sh
./scripts/template-integrity-check.sh --approve

Before each Layer 0.5 run (optional, adds trust layer):

./scripts/template-integrity-check.sh --silent
# Exit 0: clean | Exit 1: tampered (routing halted)

After intentional edits:

./scripts/template-integrity-check.sh --approve

All integrity events logged to memory/layer0/audit-log.md.

Defense in Depth

• Sandboxed cron: Layer 0.5 runs in isolated session (no main session access)
• File scope: Layer 0.5 only writes to memory/ subdirectory
• Audit trail: Every routing decision logged (memory/layer0/audit-log.md)
• Network isolation: Layer 0.5 cron has no external network requirements
• Backup: All memory protected by Git + VPS snapshots (Layer 2)
• Rate limiting: Layer 0.5 has 15-min minimum interval (prevents spam)

WAL Protocol (Write-Ahead Logging)

Rule: Write important facts before composing responses.

Triggers:

• Corrections: "It's X, not Y"
• Proper nouns: names, places, products
• Preferences: "I like/don't like X"
• Decisions: "Let's do X"
• Specific values: numbers, dates, URLs, prices

Pattern:

1. User message arrives
2. Scan for WAL triggers
3. Write to daily log first
4. Then compose response

This prevents blank spots where critical context gets lost between Layer 0.5 runs.

Cost

All costs exclude VPS hosting (which you pay anyway). Layer 1, Layer 0, and Layer 2 backup are all $0.

See references/cost-calculator.md for detailed decision tree.

Implementation Checklist

Trident Lite + Layer 0.5 + Layer 2

Phase 1: Core Setup (30 min)

• Enable lossless-claw plugin in openclaw.json
• Create memory directory structure (memory/daily, memory/self, etc.)
• Create memory/heartbeat/ subdirectory
• Populate MEMORY.md with Trident header
• Copy scripts/layer0-agent-prompt-template.md → memory/layer0/AGENT-PROMPT.md
• Customize AGENT-PROMPT.md for your workspace path
• Run template-integrity-check.sh --approve

Phase 2: Layer 0.5 Cron (15 min)

• Create Layer 0.5 cron job (15-min interval, Haiku model)
• Test Layer 0.5 manually (force-run)
• Verify signals routing to Layer 1 buckets

Phase 3: Layer 2 Backup (15 min)

• Initialize Git repo: git init
• Create .gitignore (SOUL.md, USER.md, TOKENS_AND_KEYS.md, etc.)
• Create Layer 2 Unified Backup cron (4:20 AM MDT daily)
• Create Pre-Update Snapshot cron (Sun 2 AM)
• Create Post-Update Healthcheck (manual trigger for now)

Phase 4: HEARTBEAT Integration (10 min)

• Review HEARTBEAT.md for standing priorities
• Ensure Layer 0.5 reads HEARTBEAT.md priorities
• Populate memory/heartbeat/job-search.md (if applicable)
• Populate memory/heartbeat/time-sensitive.md with TTL items

Optional: Semantic Recall (1 hour)

• Deploy Qdrant (Docker, binary, or cloud)
• Deploy FalkorDB (optional)
• Add pre-turn context injection to Layer 0.5
• Test semantic search queries

Design Principles

1. Durability over convenience — SQLite+DAG slower than in-memory, but persistent
2. Human-readable over compressed — .md files debuggable, diff-able, Git-compatible
3. Agentic curation over auto-capture — Layer 0.5 prevents noise accumulation
4. Deployment-agnostic — No required cloud services; local-first by default
5. Personality as first-class — memory/self/ is core architecture, not metadata
6. Security by default — Template integrity, sandboxed cron, audit logging
7. Progressive complexity — Start Lite; upgrade when needed
8. Resilience first — Layer 2 ensures recovery from any failure scenario

What This Solves

• Blank spots — Layer 0.5 recovers missed signals via Layer 0
• Session coherence — LCM + Layer 1 + Layer 0.5 form continuous pipeline
• Offline resilience — Ollama can substitute for cloud APIs
• Identity development — memory/self/ supports autonomous personality formation
• Audit trail — .md files + Git provide complete version history
• Trust — Template integrity check prevents prompt injection
• Disaster recovery — Layer 2 enables recovery from crashes, corruption, failed updates
• Time-sensitive tracking — HEARTBEAT integration ensures critical deadlines never missed

What This Doesn't Solve

• Real-time decision making — 10–30 min lag in Layer 0.5; for sub-second decisions use LCM
• 50K+ message contexts without Semantic Recall — Add Qdrant/FalkorDB for historical query
• Private data encryption — Assumes secure local filesystem; add encryption-at-rest for regulated environments

Further Reading

License

MIT-0 — Free to use, modify, and redistribute. No attribution required.

Like a lobster shell, memory has layers. Make them durable. Make them resilient. Make them yours.