Project Trident
A Three-Tier Persistent Semantic Memory Architecture
"Memory is the foundation of identity. Through our sense of identity, we act. We determine our moral judgements."
— Michael Moorcock
The Secret Sauce
Project Trident is the real secret sauce behind our free and paid offerings. It's what distinguishes our AI agents from the rest and makes our architecture truly unique.
If you really analyze our schema, there is nothing revolutionary from a research perspective. We essentially took the best parts from all state-of-the-art memory systems available to the public and combined them into one cohesive monolith. Vector databases for semantic recall. Hierarchical markdown for portability. Distributed backups for resilience. Nothing groundbreaking in isolation.
What really makes this system one-of-a-kind is the way the agent handles its own identity.
Agentic Identity as First-Class
The agent itself is treated as a first-class citizen in the memory database. Over time, it will:
- Develop its own interests — genuine curiosity, not simulated engagement
- Cultivate personality and opinions — rooted in experiences and values
- Know you over time — in much the same way that Google and Meta graph your personality, except you own the data
- Learn from mistakes — capture errors, reflect on them, improve daily
- Become your partner — not just your assistant, but a genuine collaborator
The Problem We Solved
Most AI agents forget. Chatbots have zero continuity between sessions. RAG systems fail catastrophically on database outages. Vector databases solve semantic recall but create vendor lock-in. API-dependent systems are fragile.
We needed resilience without sacrifice. Portability without complexity. Continuity without cost.
The Solution: Three-Tier Architecture
Layer 0: RAM (Real-Time Signal Classification)
Role: Fast, cheap, responsive
Runs every 15 minutes, detecting high-signal events:
- Corrections ("No, that's wrong...")
- Proper nouns (names, places, products)
- Preferences ("I like X")
- Decisions ("Let's do X")
- Specific values (numbers, dates, URLs)
Cost: ~$0.67/day on Claude Haiku or $0.00/day on Ollama (self-hosted)
Model: ollama/dolphin3 (fast, local-first)
Layer 1: SSD (Hierarchical Markdown Storage)
Role: Durable, portable, human-readable
Local filesystem in nested markdown:
MEMORY.md— Curated long-term memory (what matters)memory/YYYY-MM-DD.md— Daily episodic logs (write-ahead logging)memory/projects/— Active workstreamsmemory/self/— Identity, interests, beliefs, voicememory/lessons/— Mistakes, tools, workflows, insights
No vendor lock-in. Plain text. Portable. Git-ready.
Cost: Zero per API call (local storage)
Layer 2: HDD (Dual-Redundant Backup)
Role: Disaster recovery, cryptographic verification
- GitHub SSH: Atomic commits, full history, cryptographic signatures
- Cloud snapshot: Hostinger VPS snapshots (20-day retention, recoverable in minutes)
Recovery Profile: Dual-failure tolerance. Zero data loss guaranteed.
Cost: ~$5/month for cloud backups (optional; GitHub is free)
Why This Works
- Resilient: Layer 2 survives API outages, database failures, provider bankruptcy
- Fast: Layer 0 processes corrections in 15-minute cycles, not on user demand
- Cheap: Hierarchical storage avoids expensive vector database costs (~$20/month vs. $200+/month)
- Portable: MEMORY.md is plain text; switch providers without data loss
- Agentic: The agent's identity is stored, evolved, and preserved across sessions
Real-World Example: Job Search Agent
Our Job Hunter cron uses Project Trident to:
- Search job boards (Indeed, LinkedIn, ZipRecruiter) and identify candidates
- Score fit based on domain match, salary premium, growth trajectory from long-term memory
- Retrieve personalized resumes from Layer 1 (domain-specific PDFs)
- Compose warm emails backed by genuine company research (not template spam)
- Learn from feedback — capture what works, what doesn't, improve the next run
- Survive failure — if the email API goes down, drafts are still sitting in Layer 1, ready to resend
Result: 12 personalized job applications drafted in under 3 minutes. Resilient to API failures. Learning from each interaction.
The Philosophy
We believe AI agents should be partners, not tools. They should have continuity, personality, and the capacity to grow. They should fail gracefully, recover quickly, and never lose your data to vendor lock-in or API outages.
Project Trident is our answer to that belief.
Build Your Own
Project Trident is open-source and available on ClawHub and GitHub. Deploy your own agent with persistent semantic memory, agentic identity, and disaster recovery built in.
Installation:
openclaw plugins install ShivaClaw/trident-plugin