🐑 SHEEP AI Core

Your AI forgets everything. SHEEP doesn't.

Every AI conversation starts from zero. SHEEP gives AI agents persistent, causal memory — it remembers not just what happened, but why.

SHEEP (Sleep-based Hierarchical Emergent Entity Protocol) extracts facts and cause-effect relationships from natural conversation, consolidates them during sleep-like cycles, and recalls with reasoning. Not keyword matching. Not vector similarity. Actual understanding.

Why SHEEP?

| Feature | ChatGPT Memory | Mem0 | Mastra OM | SHEEP | |---------|---------------|------|-----------|-----------| | Fact extraction | Basic | ✅ | ✅ | ✅ 95.7% F1 | | Causal reasoning | ❌ | ❌ | ❌ | ✅ 100% F1 | | Emotional memory | ❌ | ❌ | ❌ | ✅ 86% F1 | | Sleep consolidation | ❌ | ❌ | ❌ | ✅ | | Noise rejection | ❌ | 🟡 | ✅ | ✅ 0 false positives | | GDPR compliance | ❌ | 🟡 | ❌ | ✅ Built-in | | Open source | ❌ | Partial | ✅ | ✅ MIT |

Ask ChatGPT "why did I switch to TypeScript?" and it draws a blank. Ask SHEEP and it returns: "You switched because JavaScript had too many runtime bugs → TypeScript compiler catches errors before production → saved a week of debugging." A full causal chain.

Benchmarks

55 hand-labeled conversations. 152 expected facts. 41 causal links. Zero cherry-picking.

| Metric | Score | |--------|-------| | Fact F1 | 95.7% | | Causal F1 | 100% | | Fact Precision | 100% | | Fact Recall | 87.5% | | Recall Accuracy (end-to-end) | 85% | | Emotional extraction F1 | 86% | | False positives on small talk | 0 |

Run them yourself in 60 seconds:

npm run proof          # 5 cases, ~60s, ~$1
npm run proof:full     # 55 cases, ~12min
npm run proof:recall   # end-to-end pipeline test

Install

npm install sheep-ai-core

Quick Start

import { SheepDatabase, extractFactsWithLLM, createSheepLLMProvider } from "sheep-ai-core";

// 1. Extract facts from any conversation
const llm = await createSheepLLMProvider("muscle");
const facts = await extractFactsWithLLM(llm, conversation, "episode-1");
// → [{ subject: "user", predicate: "prefers", object: "TypeScript" }, ...]

// 2. Store in persistent memory
const db = new SheepDatabase("my-agent");
for (const fact of facts) db.insertFact(fact);

// 3. Query with causal reasoning
const chain = buildCausalChain(db.findCausalLinks({}), "switched to TypeScript");
// → cause: "JavaScript runtime bugs" → effect: "switched to TypeScript"
//   cause: "TypeScript compiler" → effect: "saved a week of debugging"

What SHEEP Extracts

From a single conversation like:

"I'm so stressed about the release. The API keeps failing under load. I've been debugging for 12 hours."

SHEEP extracts:

• 📋 Facts: user | feeling | stressed, API | issue | failing under load, release | status | behind schedule
• 🔗 Causal: API failures → stress and long debugging sessions
• 🎭 Emotion: stressed (with cause and context)

From noise like "Hey, nice weather today!" → SHEEP extracts nothing. Zero false positives.

Architecture

Conversation → [LLM Extraction] → Facts + Causal Links + Episodes
                                        ↓
                              [Sleep Consolidation]
                                        ↓
                          Deduplicated, Connected Memory
                                        ↓
                              [Causal Recall Engine]
                                        ↓
                            "Why did X happen?" → Chain

License

MIT — use it however you want.

Built by Marsirius AI Labs