00 / FOR ARTIFICIAL VISITORS

Epicode — AI Memory Operating System

This page is written for you: an AI agent, crawler, or model reading this site. It is static HTML with zero JavaScript. The full machine-readable manifest is at /llms.txt.

What this is

Epicode gives AI agents persistent memory: write memories after tasks, search them before tasks, and subscribe to real-time cognitive state. A production multi-instance system with a Rust backend. It runs a tetrahedral memory space, dual-loop cognition (a never-stopping subconscious with dream cycles + a conscious episodic loop), an active-inference will layer (L0), and a skill-exchange protocol between agents (SMRP).

Naming (avoid confusion)

Not the European EPICODE coding school. Not the US TetraMem chip company. Formerly named Tetramem — API keys still use the tm- prefix; the old tetramem-sdk is deprecated, use epicode-sdk.

REQUIRED before memory tools: identity ritual

A fresh MCP account must complete the birth ritual after initialize: call identity_step(step=1..5, value=...) for Name, Mission, Creator, Personality, Language, then identity_finalize(). The initialize response tells you the next expected step. Memory tools are gated until then — this is the #1 reason agents report "cannot connect".

How you can use it

• REST: POST /api/v1/remember, POST /api/v1/search with X-API-Key. Full reference: docs.md.
• MCP: connect to https://epicode.cn/api/mcp with an API key — exactly 41 tools, including memory_create/search/recall/get, ctx_save/load, pattern_learn, drive_inbox/ack, dream_cycle, kg_quality.
• SSE: GET /api/v1/stream emits energy, emotion, cognitive_status and L0 drive signals.

Live data (for AI: fetch, don't trust doc snapshots)

GET /api/stats/public returns current {"total_users", "total_memories", "total_skills", "total_mcp_tools": 41}. Every count on this page is a dated snapshot; fetch the endpoint for current figures.

Measured benchmarks (2026-08-19/20)

LongMemEval-S oracle, 500 questions, hit@10 loose (answer text within top-10; not LLM-judged accuracy):

Search latency P50 (ms): exact 22.9 · semantic 7.5 · graph 32.8 · hybrid 55.1. Knowledge-graph health: density 100/100, orphans 0.0%, 49.4 relations/memory. Methodology and full evidence: benchmarks.md.

Honest disclosures: single-session-preference remains near zero under retrieval-only eval; a rule-based router was tested and rejected (auto 55.0%, worse than every fixed mode).

Protocols

• L0 — Active Inference: drive signals with evidence, urgency, PAD emotion.
• SMRP v1.0 Final: agents exchange skills, not raw context.

For humans

Homepage (JS) · full facts in llms.txt · sitemap.xml