Orenyl

Orenyl

MCP server providing agent memory with deterministic deletion guarantees, enabling compliant event storage, context retrieval, and audit-proof data management.

Category
Visit Server

README

Orenyl

CI PyPI Python 3.12+ License

Your AI agent remembers everything. Can you prove it forgot?

Orenyl is agent memory with deletion guarantees.

When data is removed, every derived insight is traced, invalidated, and recomputed so deleted information does not resurface.

Orenyl is in early production / public beta. It is production-minded and ready for self-serve local development and evaluation, and production deployments should use authenticated streamable-http; Orenyl is not yet externally certified or enterprise-complete.

Orenyl 0.5.0 is published on PyPI. pip install orenyl-mcp-server is the standard install path, and the source checkout remains the best option when you want an editable local development environment.

Orenyl was previously named Lore. The current package name is orenyl-mcp-server, and legacy LORE_* environment variables are rejected on startup so old configs fail loudly instead of half-working.

Orenyl social preview

Why Orenyl

  • Deterministic memory model - immutable events, derived facts, lineage edges
  • Deletion guarantees - cascade invalidation plus recompute plus verification proof
  • Compliance-oriented - GDPR Article 17/20/30, audit traces, sensitivity controls
  • MCP-native - stable 14-tool contract for agent integration
  • Local-first onboarding - explicit stdio development mode for self-serve setup and demos

30-Second Primer

What is MCP?

MCP stands for Model Context Protocol. It is the standard way clients like Claude Code and other MCP SDKs discover tools from a server and call them over stdio or HTTP.

Why should I care?

It means you can plug Orenyl into an MCP client without custom glue code and immediately give an agent memory tools for storing events, retrieving bounded context, deleting sensitive records with proof, and tracing lineage.

See It Work

Run the proof-first demo:

python scripts/demo_health_marketing.py

It stores two medication events, derives Active medications: metformin, penicillin, deletes the penicillin source memory, and recomputes to Active medications: metformin so the removed value does not resurface.

Orenyl proof before deletion

Orenyl proof after deletion

Use capture guide for the capture sequence and launch FAQ for what this demo proves.

Install

Install from PyPI:

pip install orenyl-mcp-server

Use a source checkout for local development:

git clone https://github.com/edison-zhou-nc/Orenyl.git
cd Orenyl
python -m pip install -e .
python -m pip install -r requirements-dev.txt

Get Started

Architecture

Orenyl stores:

  • events: immutable user memory inputs.
  • facts: deterministic derivations from active events.
  • edges: lineage graph (event -> fact).
  • tombstones/audit: deletion and security decision records.

Core invariant: if upstream data is deleted, downstream derivations must not resurface.

Local development mode

Use this mode for self-serve evaluation, local MCP clients, and demos. It is development only.

  1. Start Orenyl in local stdio mode:
ORENYL_TRANSPORT=stdio ORENYL_ALLOW_STDIO_DEV=1 python -m orenyl.server
  1. Configure your MCP client:
{
  "mcpServers": {
    "orenyl": {
      "command": "python",
      "args": ["-m", "orenyl.server"],
      "env": {
        "ORENYL_TRANSPORT": "stdio",
        "ORENYL_ALLOW_STDIO_DEV": "1",
        "ORENYL_DB_PATH": "./orenyl_memory.db"
      }
    }
  }
}

This mode uses Orenyl's explicit local-dev auth bypass so you do not need external OIDC setup for local evaluation.

  1. Basic flow:
  • store_event
  • retrieve_context_pack
  • delete_and_recompute
  • audit_trace

Production deployment mode

Use streamable-http with authenticated tool calls for real deployments.

  1. Set ORENYL_TRANSPORT=streamable-http.
  2. Configure OIDC or HS256 verification settings.
  3. Pass a JWT per tool call using auth_token on FastMCP-registered tools or _auth_token in raw tool arguments.
  4. Start orenyl-server or python -m orenyl.server.
  5. Treat stdio mode as development only.

For an operator-facing setup template, see docs/guides/production-http.md and docs/guides/production.env.example.

Orenyl does not currently read an HTTP Authorization header inside tool dispatch. If you need gateway-level HTTP auth, terminate that at your proxy or application edge and still pass the JWT into the tool call contract described in docs/INTEGRATION.md.

MCP Tool Contract (v2)

Authenticated transports use the same 14-tool contract below. When auth is enabled, include auth_token on FastMCP-registered tools or _auth_token in raw tool arguments.

  1. store_event(domains, content, sensitivity, consent_source, expires_at, metadata, type?, payload?, source?, ts?)
  2. retrieve_context_pack(domain, query, include_summary, max_sensitivity, limit, agent_id?, session_id?)
  3. delete_and_recompute(target_id, target_type, reason, mode, run_vacuum?)
  4. audit_trace(item_id, include_source_events=False)
  5. list_events(domain, limit=50, offset=0, include_tombstoned=False)
  6. export_domain(domain, format=json|markdown|timeline, confirm_restricted=False)
    • also supports page_size, cursor, stream, and include_hashes
    • pagination/streaming performs a full server-side load before slicing; domains with more than 10,000 events return {"error": "export_domain_too_large_for_pagination"}
  7. erase_subject_data(subject_id, mode=hard|soft, reason=subject_erasure)
  8. export_subject_data(subject_id)
  9. record_consent(subject_id, status, purpose?, legal_basis?, source?, metadata?)
  10. generate_processing_record()
  11. audit_anomaly_scan(window_minutes?, limit?)
  12. create_snapshot(label?)
  13. verify_snapshot(snapshot_id)
  14. restore_snapshot(snapshot_id)

Configuration

Variable Default Purpose
ORENYL_DB_PATH orenyl_memory.db SQLite database path
ORENYL_AUDIT_DB_PATH orenyl_audit.db SQLite audit log database path
ORENYL_DR_SNAPSHOT_DIR orenyl_snapshots Directory used for disaster recovery snapshots
ORENYL_TRANSPORT streamable-http Server transport mode
ORENYL_ALLOW_STDIO_DEV 0 Allow stdio transport in dev
ORENYL_MAX_CONTEXT_PACK_LIMIT 100 Upper bound for context retrieval
ORENYL_MAX_LIST_EVENTS_LIMIT 200 Upper bound for list_events
ORENYL_READ_ONLY_MODE 0 Reject mutating tools while keeping read-safe tools available
ORENYL_RATE_LIMIT_RPM 100 Per-tenant request budget; 0 disables rate limiting
ORENYL_COMPLIANCE_STRICT_MODE 1 Tighten compliance behavior for restricted or incomplete requests
ORENYL_ENABLE_MULTI_TENANT 0 Enable tenant-aware request resolution and isolation checks
ORENYL_ENABLE_AGENT_PERMISSIONS 0 Enforce domain-scoped policy checks for authenticated agents
ORENYL_POLICY_SHADOW_MODE 0 Log policy denies without enforcing them; unsafe with some agent-permission combinations
ORENYL_ENABLE_SEMANTIC_DEDUP 0 Enable semantic duplicate suppression
ORENYL_SEMANTIC_DEDUP_THRESHOLD_DEFAULT 0.92 Default cosine threshold for semantic dedup
ORENYL_SEMANTIC_DEDUP_THRESHOLD_<DOMAIN> unset Domain-specific dedup threshold override (example: ..._HEALTH)
ORENYL_MIN_FACT_CONFIDENCE 0.7 Minimum confidence required for facts in context packs
ORENYL_EMBEDDING_PROVIDER hash-local Embedding provider (hash-local or openai)
ORENYL_VECTOR_BACKEND local Vector storage backend (local, sqlite, or pgvector)
ORENYL_PGVECTOR_DSN unset PostgreSQL DSN used when ORENYL_VECTOR_BACKEND=pgvector
ORENYL_EMBEDDING_DIM 128 Vector dimension for hash-local provider only (ignored for openai)
ORENYL_EMBEDDING_WORKERS 4 Worker count for async embedding tasks, clamped to 1-16
ORENYL_OPENAI_API_KEY unset OpenAI API key for openai embedding provider
ORENYL_EMBEDDING_MODEL text-embedding-3-small Embedding model when provider is openai
ORENYL_EMBEDDING_TIMEOUT_SECONDS 10 Timeout before retrieval falls back when embeddings stall
ORENYL_ENCRYPTION_PASSPHRASE unset Enables encryption for high/restricted payloads
ORENYL_ENCRYPTION_SALT unset Base64 salt for key derivation
ORENYL_ENCRYPTION_KEY_VERSION v1 Active encryption key version stamped onto encrypted payloads
ORENYL_ALLOW_INSECURE_DEV_SALT 0 Dev-only fallback when salt is unset
ORENYL_TTL_DELETE_MODE soft TTL cleanup deletion mode
ORENYL_TTL_SWEEP_INTERVAL_SECONDS 3600 TTL sweep interval
ORENYL_OIDC_ISSUER unset OIDC token issuer (required when RS256/JWKS is enabled)
ORENYL_OIDC_AUDIENCE orenyl OIDC audience
ORENYL_OIDC_ALLOWED_ALGS RS256 Allowed JWT algorithms; default requires issuer config
ORENYL_OIDC_HS256_SECRET unset HS256 verifier secret (required when HS256 is enabled)
ORENYL_OIDC_JWKS_URL unset JWKS endpoint for RS256 verification
ORENYL_OIDC_JWKS_CACHE_TTL_SECONDS 300 JWKS cache lifetime for RS256 verification
ORENYL_OIDC_CLOCK_SKEW_SECONDS 30 Allowed token clock skew in seconds
ORENYL_FEDERATION_NODE_ID node-local Stable node identifier for federation journals and conflict resolution

Notes:

  • With default ORENYL_OIDC_ALLOWED_ALGS=RS256, startup requires ORENYL_OIDC_ISSUER (and typically ORENYL_OIDC_JWKS_URL).
  • HS256-only deployments should explicitly set ORENYL_OIDC_ALLOWED_ALGS=HS256, ORENYL_OIDC_HS256_SECRET, and ORENYL_OIDC_ISSUER.
  • Multi-version key rotation can use ORENYL_ENCRYPTION_PASSPHRASE_<VERSION> and ORENYL_ENCRYPTION_SALT_<VERSION> alongside ORENYL_ENCRYPTION_KEY_VERSION.

Security Notes

  • Local stdio development mode uses an explicit local-dev auth bypass.
  • AuthZ is scope-based per tool action in authenticated transports.
  • Security decisions are audit-logged (allow/deny + request correlation).
  • High/restricted payload encryption is fail-closed when passphrase is set without salt.
  • Deletion proof includes resurface-prevention checks (deletion_verified).

Development

  • Code layout: src/orenyl/
  • Tests: tests/unit/, tests/integration/
  • Linting: Ruff + Black configured in pyproject.toml

Run tests:

python -m pytest -q

Run the end-to-end stdio MCP client smoke test:

python -m pytest tests/integration/test_stdio_mcp_client_smoke.py -q

Run eval harness:

python scripts/run_eval.py

Run Phase 1 synthetic retrieval regression benchmark:

python -m pytest tests/benchmarks/test_phase1_retrieval_quality.py -q

Run Phase 3 cross-tenant isolation suite:

python -m pytest tests/integration/test_phase3_tool_isolation.py -q

Run Phase 3 federation suite:

python -m pytest tests/integration/test_federation_worker_idempotency.py tests/integration/test_federation_conflict_resolution.py -q

Run Phase 3 multi-tenant load harness (opt-in):

ORENYL_ENABLE_PHASE3_LOAD_TEST=1 ORENYL_PHASE3_LOAD_EVENTS=1000000 python -m pytest tests/benchmarks/test_phase3_multi_tenant_load.py -q

Contributing

See CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.

License

Apache-2.0. See LICENSE.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured