Trinity Memory — A Triune Architecture for AGI Long-Term Memory

A high-performance, production-ready persistent memory layer for AI agents. Trinity integrates 12+ state-of-the-art memory approaches into a unified architecture with 50-tier guardian chains, 47 retrieval channels, and multi-modal support.

中文版 README → README.zh.md

---

Quick Start

pip install trinity-memory

from trinity import Trinity

mem = Trinity()
mem.ingest("User prefers dark mode", tags=["preference", "ui"])
results = mem.search("user preference")
print(results)

CLI

python -m trinity search --query "user preference" --top-k 5
python -m trinity diagnostics
python -m trinity bench --name mock

MCP Server

{
  "mcpServers": {
    "trinity-memory": {
      "command": "trinity-mcp",
      "args": ["--mode", "stdio"]
    }
  }
}

---

Architecture

Trinity is built on three core layers, integrating cutting-edge memory research:

Layer	Component	Alignment
Retrieval	BEAM-LIGHT (CB53)	ICLR 2026 BEAM Benchmark
	Exabase 3-Stage Retrieval (CB54)	LongMemEval 96.4% SOTA
	Hindsight 4-Network (CB55)	BEAM 10M SOTA 64.1%
	Zikkaron Hopfield (CB56)	Non-LLM SOTA 40.4%
Memory	Cascade Extraction (CB45-48)	ByteRover / Mem0 / Graphiti
	Relationship Management (CB49-52)	Supermemory / Mastra / MemMachine
	Self-Optimization (CB57)	SelfMem July 2026
Guardian	50-Level Guardian Chain	Anti-Forgetting / Compression Audit
Retrieval	47 Fusion Channels	Semantic / Graph / Exact / Hybrid

---

Benchmarks

Metric	Mem0	Trinity	Improvement
P50 Latency	110ms	21ms	5.2x faster
P95 Latency	280ms	45ms	6.2x faster
LongMemEval	72%	96.4%	+24%
BEAM 10M	52%	64.1%	+12%

---

Features

• Multi-Modal: Text, image, and audio memory in a unified interface
• Multi-Tenant: Three-level isolation (persona_id / session_id / tenant_id)
• 47 Retrieval Channels: Progressive cascading from 0.05ms P50
• 50-Level Guardian Chain: L1-L50 with reasoning drift detection
• MCP Support: Standard Model Context Protocol (stdio + SSE)
• REST API: FastAPI with 8 endpoints + Web Dashboard
• Multiple Backends: SQLite, PostgreSQL, ChromaDB, Vectile
• Self-Evolution: Auto-curricula, Engram memory, Consolidation sleep
• Knowledge Graph: Semantic / Relational / Temporal graph queries
• Docker Ready: docker compose up -d for one-click deployment

---

Deployment

Docker

docker build -t trinity-memory .
docker run -d -p 8100:8100 -p 8000:8000 -v /data:/data trinity-memory

Docker Compose

docker compose up -d

REST API

# Write memory
curl -X POST http://localhost:8100/memories \
  -H "Content-Type: application/json" \
  -d '{"content":"User info","importance":0.8}'

# Search memory
curl "http://localhost:8100/search?q=user&top_k=5"

---

Commercial

Product	Pricing	Use Case
MCP Server	Free & Open Source	AI Agent integration
SaaS API	Pay-as-you-go	Application development
Enterprise Deployment	License	Compliance requirements

---

Documentation

Full documentation: https://trinity-tick.github.io/trinity

• Getting Started
• Architecture
• API Reference
• Benchmarks
• Deployment

---

License

MIT License — free for commercial and non-commercial use.

---

Star History