pce-memory
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@pce-memoryfind memories related to the design meeting"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
PCE Memory
Process-Context Engine (PCE) - MCP Server for Context-Aware Memory & Retrieval
Overview
PCE Memoryは、LLMエージェントやアプリケーションに文脈記憶・検索・統合機能を提供するMCPサーバーです。
Key Features
🔒 Boundary-first Security: スコープベースの境界管理とRedact-before-send
🧠 Dual-phase Memory: LCP (長期記憶) + AC (短期記憶) の二相アーキテクチャ
🔍 Hybrid Search: FTS + VSS のハイブリッド検索 (r, S, g 関数)
📊 DuckDB Storage: 組み込みDuckDBによる高速ローカルストレージ
🎯 MCP Protocol: Model Context Protocol v1.0.4 準拠
🛡️ Law-Driven Engineering: fp-ts/io-ts によるType-safe & Property-based Testing
Related MCP server: mcp-synaptic
Benchmark Results
pce-memory's hybrid search pipeline is evaluated across three dimensions: search quality, scalability, and latency.
Search Quality (Ablation Study)
Search Method | Recall@10 | nDCG | Latency |
Text-only (BM25) | 50.0% | 50.0% | 34.7ms |
Vector-only (Semantic) | 87.1% | 69.0% | 32.3ms |
Hybrid (BM25 + Vector) | 91.7% | 81.0% | 32.6ms |
Hybrid search combines the precision of keyword matching with the semantic understanding of vector search, achieving +41.7pp recall over text-only and +4.6pp over vector-only.
Provenance-Aware Reranking
The g() reranking function leverages claim provenance quality to improve result ordering:
Metric | Without Rerank | With Rerank | Delta |
Recall@10 | 65.2% | 74.2% | +9.1pp |
nDCG | 62.4% | 67.1% | +4.7pp |
Measured at 150 claims (golden + synthetic noise). Claims with richer provenance (actor, notes) are ranked higher.
Scalability
Claims | P@5 | R@10 | MRR | Latency p50 |
15 | 18.2% | 91.7% | 83.0% | 38ms |
50 | 18.2% | 87.9% | 79.8% | 42ms |
100 | 17.3% | 87.9% | 78.9% | 40ms |
250 | 17.3% | 81.1% | 77.7% | 43ms |
500 | 15.5% | 72.0% | 68.6% | 45ms |
1,000 | 15.5% | 72.0% | 68.6% | 57ms |
5,000 | 12.7% | 60.6% | 61.4% | 60ms |
Latency Profile
Operation | Time |
Embedding model cold start | 177ms (once per session) |
Embedding (cached) | 0.1ms |
Search p50 (with rerank) | 37.5ms |
Rerank overhead | 5.1ms |
All operations are well below the 100ms human perception threshold. Run pnpm benchmark to reproduce (requires apps/pce-memory/external/assay-kit submodule).
Project Structure
pce-memory/
├── apps/
│ └── pce-memory/ # MCP server implementation
├── packages/
│ ├── pce-boundary/ # Boundary validation & redaction
│ ├── pce-embeddings/ # Embedding provider abstraction
│ ├── pce-policy-schemas/ # YAML policy schemas
│ └── pce-sdk-ts/ # TypeScript client SDK
├── docs/ # Documentation
│ ├── pce-memory-vision.md
│ ├── core-types.md
│ ├── mcp-tools.md
│ ├── boundary-policy.md
│ └── activation-ranking.md
├── scripts/ # Development and local validation helpers
└── validation/ # Local validation tasks and result artifactsQuick Start
Prerequisites
Node.js: ≥ 22.18.0 (LTS)
pnpm: ≥ 9.0.0
Installation
# Install dependencies
pnpm install
# Build all packages
pnpm build
# Run tests
pnpm test
# Run property-based tests
pnpm test:pbtRunning the MCP Server
# Start MCP server (stdio transport)
cd apps/pce-memory
pnpm devDevelopment Workflow
# Watch mode (auto-reload on file changes)
pnpm dev
# Type checking
pnpm typecheck
# Linting
pnpm lint
pnpm lint:fix
# Formatting
pnpm format
pnpm format:check
# Clean build artifacts
pnpm cleanLocal Architecture Validation with Ollama
The repository includes a documented local validation workflow for architecture experiments using Ollama and qwen3.5:122b-a10b.
# Interactive Codex against Ollama
pnpm local:codex
# Interactive Claude Code through Ollama launch integration
pnpm local:claude
# Canonical smoke task
pnpm local:validation:smokeSee docs/local-validation-ollama.md for machine assumptions, launch recipes, known limitations, and result capture conventions.
MCP Tools
PCE Memoryは以下のMCPツールを提供します:
Tool | Description |
| ポリシー適用と状態初期化 |
| raw observation を記録(短期TTL、durable化はしない) |
| observation / claim / active context から昇格候補を作成 |
| 候補を durable claim に昇格 |
| durable claim の append-only repair |
| 既に蒸留済みの durable knowledge を直接登録 |
| intent-aware に Active Context を構成 |
| 境界チェック / redact-before-send |
| durable claim へのフィードバックを送信 |
| 状態情報を取得 (state/policy_version) |
|
|
詳細は docs/mcp-tools.md を参照してください。
Recommended V2 Flow
現在の推奨フローは次のとおりです。
observe -> distill -> promote -> activate(intent) -> feedback -> rollbackobserveは raw-only です。durable claim は inline で作りません。upsertは already-distilled な durable knowledge の escape hatch です。scope=sessionとboundary_class=secretの durable 書き込みは避け、observeを使います。
Observation(pce_memory_observe)の保持とセキュリティ(要点)
Observation は短期TTLで保持し、期限後は
contentをスクラブ(NULL化)する運用を推奨します。PCE_OBS_TTL_DAYS/PCE_OBS_TTL_DAYS_MAXで TTL を調整できます。PCE_OBS_MAX_BYTESでcontentの最大バイト数を制限できます(既定 65536)。PCE_OBS_STORE_MODE(raw|redact|digest_only)で保存モードを調整できます(既定redact)。secret を検知した場合は fail-safe として
contentを保存せず、抽出もスキップします(詳細はdocs/mcp-tools.mdの observe を参照)。
Architecture
PCE Memoryは以下のコアモジュールで構成されます:
Observation Store: raw observation を TTL 付きで保持し、必要に応じて redact / digest-only を適用
Promotion Pipeline:
distill -> promote -> rollbackで durable claim を管理Retriever:
activateで intent-aware に文脈を検索・活性化 (Query + LCP -> AC)Critic: durable claim を評価・更新 (Feedback -> utility/confidence/recency)
詳細は docs/pce-memory-vision.md を参照してください。
Testing Strategy
Law-Driven Engineering (LDE)
PCE Memoryは**Law-Driven Engineering (LDE)**原則に従います:
Type-driven Design: fp-ts/io-ts によるBranded Types
Property-based Testing: fast-checkによる不変条件検証
Errors as Values: Either/TaskEither による明示的エラー処理
Detroit School TDD: 実際のコンポーネント連携をテスト
Formal Verification
pnpm formal:tla— TLA+ (TLC)。tlcが無い場合は Dockerghcr.io/tlaplus/tlaplusを自動利用。pnpm formal:alloy— Alloy (SAT4J)。初回実行時に jar を.cache/formalに取得し Java で全コマンドを走査。pnpm formal:all— 上記まとめ実行。
Test Commands
# All tests (unit + property-based)
pnpm test
# Property-based tests only
pnpm test:pbt
# With coverage (≥ 80% required)
pnpm test --coverage
# Watch mode
pnpm test:watch詳細は packages/pce-boundary/test/README.md を参照してください。
Contributing
Code Quality Gates
Gate | Threshold |
Static Analysis | 0 errors |
Coverage | ≥ 80% |
Cyclomatic Complexity | ≤ 10 |
Mutation Score | ≥ 60% (optional) |
Commit Convention
feat: Add new feature
fix: Fix bug
docs: Update documentation
test: Add tests
refactor: Refactor code
chore: Update build configLicense
MIT License - see LICENSE for details
Documentation
Related Projects
Acknowledgments
Built with ❤️ by CAPHTECH, following Law-Driven Engineering principles and inspired by:
Ernst Cassirer's Philosophy of Symbolic Forms
Actor-Network Theory (ANT)
Buddhist concepts of dependent origination (縁起)
Status: 🚧 Work in Progress (MVP Phase)
This server cannot be deployed
Maintenance
Related MCP Connectors
An MCP memory server. One memory your agents share — across models, devices and apps.
Person-owned AI memory that learns, not just stores — portable context for any MCP client.
- EngramOAuthapp.getengram
Persistent, verbatim, searchable memory for AI assistants — one memory across every MCP client.
Persistent personal memory for AI assistants — save, search, and recall across every MCP client.
Related MCP Servers
- AlicenseNot gradedqualityNot gradedmaintenanceAn MCP server enabling AI assistants to store, retrieve, and manage contextual information across conversations with features like persistent memory, advanced search, tagging, and privacy controls.MIT
- AlicenseNot gradedqualityCmaintenanceA memory-enhanced MCP server with RAG database and expiring memory capabilities, enabling AI systems to store and retrieve memories and documents with semantic search.3MIT
- AlicenseNot gradedqualityDmaintenanceA local-first memory MCP server that enables storing, searching, and managing personal memories with hybrid keyword and semantic recall, all on-device.138 npmMIT
- AlicenseNot gradedqualityFmaintenanceMCP server providing cognitive memory storage with dual memory system (project and user), semantic classification, time knowledge graph, and memory decay, enabling persistent context-aware interactions.MIT