Noesis — Adaptive Epistemic Memory Engine

Wave-field memory with semantic recall, lexical reranking, and MCP protocol support.

What it is

Noesis is a memory server that learns from usage. Unlike static RAG systems, every confirm/reject signal adjusts concordance scores in the wave field — the system's knowledge converges toward truth over time.

Quick start (MCP client)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "noesis": {
      "command": "npx",
      "args": ["@vanta-ide/noesis-server", "mcp"],
      "env": {
        "NOESIS_API_KEYS": "{\"your-key\":\"your-tenant\"}",
        "EME_DATA_DIR": "/path/to/your/data"
      }
    }
  }
}

Or run directly:

npx @vanta-ide/noesis-server mcp

HTTP server

npx @vanta-ide/noesis-server

Starts on port 3100 (configurable via PORT env var).

Environment variables

| Variable | Required | Description | |---|---|---| | NOESIS_API_KEYS | Production | JSON map of API key → tenant name | | EME_DATA_DIR | Recommended | Path for SQLite database (default: /data) | | OLLAMA_URL | Optional | Ollama endpoint for semantic embeddings (default: http://localhost:11434) | | EMBED_MODEL | Optional | Embedding model name (default: nomic-embed-text) | | PORT | Optional | HTTP server port (default: 3100) | | NOESIS_ALLOW_ANON | Dev only | Set to 1 for unauthenticated local dev — never use in production |

Semantic recall

Noesis blends three signals on every recall:

• Cosine similarity — nomic-embed-text 768d vectors via Ollama
• Lexical reranker — phrase matching, term proximity, density, position bias
• Wave-field scoring — activation × concordance × durability × utility

Ollama is optional. If unreachable, recall degrades gracefully to lexical + wave-field only.

To enable semantic recall locally:

ollama pull nomic-embed-text

Domains

Eight built-in reasoning domains with per-domain decay and Synapse presets: general, music, code, finance, healthcare, research, creative, game

Healthcare domain: escalation and sync disabled by design — stays local only.

MCP tools (21)

Core: noesis_store, noesis_recall, noesis_observe, noesis_access, noesis_stats

Domain: noesis_domain_activate, noesis_domain_deactivate, noesis_domain_list, noesis_domain_health, noesis_domain_store, noesis_domain_recall, noesis_domain_confirm, noesis_domain_reject, noesis_domain_decay

Sync: noesis_sync_export

Skills: noesis_skill_propose, noesis_skill_advance, noesis_skill_evidence, noesis_skill_list, noesis_skill_promoted

Catalog: noesis_list_domains

The learning loop

1. Store knowledge with noesis_store or noesis_domain_store (include sourceRef, chunkIndex, pageNum for provenance)
2. Recall with noesis_recall — blended semantic + lexical + wave-field ranking
3. Confirm good results with noesis_domain_confirm — concordance rises
4. Reject bad results with noesis_domain_reject — concordance falls
5. System improves with each signal — static RAG doesn't do this

Docker (self-hosted with sovereign embeddings)

docker build -t noesis-server .
docker run -p 3100:3100 -v /your/data:/data \
  -e NOESIS_API_KEYS='{"sk-prod":"your-tenant"}' \
  noesis-server

Ollama and nomic-embed-text are baked into the image — no external embedding API required.

License

Proprietary — Vanta IDE. All rights reserved.