Trinity — Open Memory Layer with Governance
v8.2.0 — 治理优先的记忆操作系统:模型会换、框架会换,但记忆不换。 Trinity 是让记忆可迁移、可治理、可交易的基础设施。
Trinity is not a "memory library." It is a Memory Operating System — an infrastructure layer that any memory store (vector DB, graph DB, SQLite) can plug into, with retrieval, governance, identity, evolution, and economic protocols on top.
定位(2026-08-15, V2):记忆是 AI 最后的切换成本。Trinity = 开放记忆层 + 治理底座—— 记忆可进可出(可迁移标准)、企业敢存(治理/合规/审计)、 记忆值钱(TrustExchange 市场)。
Architecture
┌──────────────────────────────────────────────────────────┐
│ Agent Layer A2A v0.3 · DSH 原生 · 共享聚合池 · 身份 │
├──────────────────────────────────────────────────────────┤
│ Governance Layer RBAC(6) · 50-Guardian · 审计签名 · 加密 │
│ B3 策略层(isolated/shared/delegated) │
├──────────────────────────────────────────────────────────┤
│ Memory Layer 47 通道 · PPR · 意图压缩 · 蒸馏 11x │
│ 个性化(PAHF) · 跨模态 · 联邦 │
├──────────────────────────────────────────────────────────┤
│ Storage Layer SQLite(FTS5) · PostgreSQL · AES-GCM 加密 │
├──────────────────────────────────────────────────────────┤
│ Economic Layer TrustExchange 记忆市场 · 资产定价 │
└──────────────────────────────────────────────────────────┘
528 Python files · 206K+ lines · 147 API endpoints · 815 tests passing
Quick Start
# 安装(系统 Python 3.11+)
cd trinity
pip install -e .
# 验证
python -c "import trinity; print(trinity.__version__)" # → 8.2.0
# 全量测试(815 passed / 50 skipped / 0 failed,系统 Python 3.14)
python -m pytest tests/ -q
服务(全在线,supervisor 自愈)
| 服务 | 端口 | 说明 |
|---|---|---|
| trinity-api | :8001 | REST(147 端点) |
| trinity-mcp | :8000 / :8003 | MCP SSE / MCP v2 streamable-http |
| gateway | :8002 | OpenAI/Mem0 兼容层(DeepSeek 上游,鉴权/限流/模型映射) |
| dashboard | :3005 | 可视化 |
| PostgreSQL | :5430 | 维护镜像(docker) |
Key Features(名实一致,2026-08-15 实测)
| 能力 | 状态 | 说明 |
|---|---|---|
| 41 个 active 模块 | ✅ | 运行路径可达(另有 261 个论文对齐储备,status: orphan 标注,audit_modules.py 审计) |
| 47 通道检索 | ✅ | BM25+jieba / FAISS HNSW / Exabase / BEAM-LIGHT / Hindsight / PPR 图扩散 / RRF 融合 |
| 语义缓存 | ✅ | Redis 305x,scope 隔离 |
| 存储加密 | ✅ | AES-256-GCM 可选(TRINITY_STORAGE_ENCRYPTION),FTS/哈希链兼容 |
| 治理策略层 | ✅ | B3:YAML 策略(isolated/shared/delegated)+ 热切换 + 审计 |
| 多智能体 | ✅ | A2A v0.3 + 共享聚合池 + 身份漂移检测 |
| 意图压缩 | ✅ | SimpleMem 对齐(TRINITY_INTENT_CLUSTER=on) |
| 结构化蒸馏 | ✅ | ICML 2026 对齐,11x 压缩(TRINITY_DISTILL_COMPRESS=on) |
| 个性化 | ✅ | PAHF 双反馈(Meta ICLR 2026 对齐) |
| 跨模态 | ✅ | 图搜文/文搜图闭环 |
| DSH 结构融合 | ✅ | 6 表自动同步(会话/事件/goal/todo/header/schedule),goal objective 100% |
| 记忆可迁移 | ✅ | memory_portability.py:标准 JSON/NDJSON + Mem0/Zep 导入 |
| 记忆市场 | ✅ | TrustExchange:挂单/订单簿/定价/声誉(11 端点) |
| 联邦 | ✅ | 多实例 export/import/diff 同步 |
记忆可迁移(V2 核心)
# 导出标准格式(记忆护城河入场券:可进可出)
python scripts/memory_portability.py export --out memories.json
python scripts/memory_portability.py export --out memories.ndjson --format ndjson
# 导入(幂等:content_hash 去重)
python scripts/memory_portability.py import --file memories.json
# 从 Mem0 / Zep 迁移
python scripts/memory_portability.py import-mem0 --file mem0_export.json --persona p1
python scripts/memory_portability.py import-zep --file zep_export.json --persona p1
Benchmark(实测;官方 LongMemEval_S 已跑,2026-08-16)
| Benchmark | Score | 口径 |
|---|---|---|
| LongMemEval_S(官方 ICLR 2025,500 题) | session R@5 = 0.968 · turn R@5 = 0.922 · hit pos 1.3 | 官方数据集实测(hf-mirror 获取),hybrid top-5 |
| SQuAD v1.1 (adapted) | R@5 = 98.3% | 180 题 passage selection(本地) |
| LoCoMo (subset) | R@5 = 0.88 | 38 题会话聚合(中文本地集) |
| pytest | 815 passed / 0 failed | 全量 |
| LongMemEval_S 500 题 QA(judge3 三票,RouteReasoner 产品化策略路由 + pref-inner2) | 68.6%(343/500) | 2026-08-17 全量;SS-A 96.4 / SS-U 92.9 / KU 69.2 / TR 65.4 / SS-P 56.7 / MS 49.6 |
| LongMemEval_S 500 题 QA(judge3 三票,route2 benchmark 脚本) | 63.2%(316/500) | 2026-08-17 基线;MS 43.6 / SS-P 20.0 |
| LongMemEval_S 500 题 QA(dated,旧 judge) | 54.0% | 2026-08-16 全量实测 |
📊 官方 LongMemEval_S 详情与分题型:docs/bench-official/LongMemEval_S_REPORT_20260816.md QA accuracy(DeepSeek judged,官方模板,500 题)= 54.0%(dated 优化:时间戳+全量证据+ temporal 分步推理;优化前基线 49.6%,temporal-reasoning +15.7pp)。 分题型:assistant 91% / user 87% / knowledge-update 64% / multi 36% / temporal 44% / preference 3%。
⚠️ 口径声明(2026-08-16):README 旧版引用的 "LongMemEval 96.4% / BEAM 10M 64.1%" 系 Exabase M-1 / Hindsight 的成绩,非 Trinity 实测,已移除。BEAM/LoCoMo 英文官方集 仍未跑(网络限制),不构成对外宣称。
Research Foundation
Trinity 的 second_brain 与 2026 前沿对齐:PPR/HippoRAG 2、SimpleMem (ICML 2026)、
Structured Distillation (11x)、PAHF (Meta ICLR 2026)、Hindsight/BEAM、Mem0/Zep/Graphiti 思路。
Requirements
- Python 3.11+(推荐 3.14)
- Docker Desktop(容器化部署)
- jieba(中文分词)、fastapi、strawberry-graphql
License
MIT — see pyproject.toml.
Trinity: model changes, framework changes, memory doesn't.
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file trinity_memory-8.2.0.tar.gz.
File metadata
- Download URL: trinity_memory-8.2.0.tar.gz
- Upload date:
- Size: 4.2 MB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.14.5
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
27ae3311eea6822a206836195527b0ecb557b7ec7913bb382db2954645ec4829
|
|
| MD5 |
ceee20c176b56f23abab2c3a67ef2ccf
|
|
| BLAKE2b-256 |
78f765603d0c9d4a1d2486d4faa0f6e6043740f61f0eccca3a2a63e336361ede
|
File details
Details for the file trinity_memory-8.2.0-py3-none-any.whl.
File metadata
- Download URL: trinity_memory-8.2.0-py3-none-any.whl
- Upload date:
- Size: 4.1 MB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.14.5
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
cbdb91a426261b565f458037a23e71769b61cb1d98f0392580b5205efb018bb9
|
|
| MD5 |
95f9d806552d8e9beac5e48b6fdcd867
|
|
| BLAKE2b-256 |
2537227ba387166ef121c33e04cb81b86e1324839517e7df763ae6e9cbe1cd9f
|