Entelecheia — Agent Platform and Memory

Entelecheia is the agent platform: the scepter runtime that orchestrates specialized agents (the "souls"), maintains long-term memory (PhiLia), provides semantic search, and hosts the industrial integration layer. Behind the capabilities of Arona and Chest stands this platform. This guide is organized by capability: agents, memory, search, knowledge and connections.

Architecture at a glance

Clients: Arona (gateway) · Shittim Chest (chat/panels) · TUI/CLI
        │  JSON-RPC 2.0 over WebSocket (token or API key)
        ▼
   scepter runtime (node-3:8424)
   ├─ Agent manager: L1 souls (PhiLia, Skopeo, Hubris, Kalos, …)
   ├─ Skill chains: pipelines of LLM + tool calls with RAG prefetch
   ├─ PhiLia: long-term memory (vector + graph, pgvector-backed)
   ├─ ApoRia: knowledge base + workspace index (semantic search)
   ├─ OreXis: policy / safety gates on tool execution
   └─ Reflection: lesson store re-injected into prompts

1. Agents (souls)

Each soul is a specialized agent with its own identity document, tools (MCP-style), and skills. Skill chains compose LLM calls with tool execution; before each call, relevant long-term memories and knowledge-base content are prefetched and injected into the system prompt.

Safety: tool execution passes through OreXis policy gates, and industrial writes require explicit approval flows.

2. Long-term memory (PhiLia)

PhiLia is the memory service behind Arona's memory gateway:

• Store — episodes, entities and artifacts are stored as nodes in a memory graph, embedded and mirrored to pgvector (`philia_chunks`).
• Query — semantic retrieval combines vector similarity, graph traversal and recency decay (14-day half-life).
• Consolidate — periodic merging links related nodes.
• Wire surface — first-class methods Sync.MemoryStoreRequest / `MemoryQueryRequest` / `MemoryDeleteRequest` (RBAC: SystemWrite / SystemRead) alongside the generic `Mcp.CallTool` route.

Embedding: configured via OLLAMA_HOST + OLLAMA_EMBED_MODEL (e.g. nomic-embed-text), or a remote API, falling back to a local ONNX model.

3. Semantic search

Sync.SearchRequest fuses two stores into one ranked list:

• ApoRia — workspace index, agent reports and knowledge-base documents (hybrid vector + keyword with RRF).
• PhiLia — long-term memories (philia_memory source).

4. Knowledge base

Create knowledge bases, add documents and subscribe to rag subscriptions — all persisted in Postgres. Documents are embedded into the ApoRia store and retrievable through the same search surface.

5. Reflection

Lessons learned are stored in a lesson store (pgvector) and re-injected into future prompts — a second, lightweight persistent memory alongside PhiLia.

6. Connecting clients

• WebSocket ws://<host>:8424/ws — authenticate at upgrade with `?token=<connection token>` (or Bearer); then `Sync.ConnectHandshake`.
• HTTP JSON-RPC POST /api/rpc?token=… for request/response use.
• Connection token: ~/.config/entelecheia/scepter.token on the scepter node.

Env reference (subset)