scm-memory 0.7.4

SCM: memory that works like yours — wake + sleep phases for AI agents.

SCM — Memory that works like yours

Mem0 stores facts your agent saw. SCM learns from them while it's idle.

SCM (Sleep-Consolidated Memory) is the first open-source agent memory layer with both a wake phase and a sleep phase — like the only memory system in nature that actually works.

📄 Paper (35 pages) · 🛠 Deployment guide · 🔌 Integrations · 🗺 Roadmap · 📊 Benchmarks

---

What's different

Every other agent memory product (Mem0, MemGPT, A-Mem, RAG layers) does the same thing in different shapes: store facts, retrieve facts. They never think between sessions. Your agent forgets to think the moment you stop talking to it.

That's not how memory works in any system that does it well — including yours. Sleep is when memory consolidates. When you sleep, your hippocampus replays the day's experiences and your cortex abstracts patterns from them. You wake up with a better version of yesterday's understanding, not the same one.

SCM does both phases:

Phase	What happens	Bio analog	SCM modules
Wake	Selective attention. Encoding-by-importance. Cue-driven retrieval. Contradiction handling. Bounded working memory (~7 items).	Hippocampal encoding, working memory, cue-driven recall	Phases 1-5
Sleep	Pattern abstraction. Contradiction resolution. Adaptive forgetting. Knowledge-gap detection and curiosity-driven filling. Wake summary report.	NREM consolidation, REM dreaming, synaptic homeostasis	Phase 7 (M1-M6)

The result the user sees: when they come back from being away, the agent reports what it noticed.

> What did you notice while I was away?

While you were away I noticed three things:
  • You've changed jobs — I've moved you from Northstar Robotics to Atlas Labs.
  • Your Tuesday-morning runs and Friday-night dinners with Mara have become weekly patterns.
  • You've mentioned 'OAuth flow' five times without explaining it; I read up on it.
    (Authorization protocol, redirect-based, token + scope.)

This moment is the product. No other open-source memory system has it.

---

Five-line quickstart

from src.chat.engine import ChatEngine

engine = ChatEngine(profile="chatbot")
engine.chat("Hi, I'm Saish. I run every Tuesday morning.")
engine.chat("Tuesday again — out for a 5K.")
engine.force_sleep("deep")
print(engine.wake_summary().narrative)
# → "While you were away I noticed Tuesday-morning running has become a pattern..."

Or, drop the SCM MCP server into Claude Desktop / Cursor / any MCP client and add five tools (add_memory, search_memory, consolidate, wake_summary, forget) automatically. See docs/INTEGRATIONS.md.

---

Install

git clone https://github.com/Saish15/sleepai.git
cd sleepai
python -m venv venv && source venv/bin/activate
pip install -e .

For the recommended privacy-first profile (free, local, no cloud calls):

ollama pull nomic-embed-text     # 274 MB embedding model (recommended)
ollama pull llama3.2:latest      # ~4 GB chat / extraction model

That's it. SCM auto-detects Ollama and uses it. Four deployment profiles documented in docs/DEPLOYMENT.md:

Profile	LLM	Embedding	Cost / 30 turns	Privacy
A Offline-only	heuristic regex	sentence-transformers MiniLM	$0	100% local
B Ollama-only (recommended)	Ollama llama3	Ollama nomic-embed-text	$0	100% local
C Hybrid	DeepSeek-chat	Ollama nomic-embed-text	~$0.04	text→cloud, vectors local
D All-cloud	OpenAI gpt-4o-mini	OpenAI text-embedding-3-large	~$0.06	all→cloud

---

Works seamlessly with any LLM and any harness

SCM doesn't care which LLM you use. Concept extraction goes through LLMExtractor; switch providers with one env var:

LLM_PROVIDER=ollama       # local, free
LLM_PROVIDER=deepseek     # cheap cloud
LLM_PROVIDER=openai       # premium
# Anthropic / Voyage / Together / any OpenAI-compatible endpoint also works

Same for embeddings — sentence-transformers, Ollama, or any OpenAI-compatible provider.

Drop SCM behind any agent framework:

• MCP server (Claude Desktop, Cursor, ChatGPT-with-MCP) — scm mcp in your config
• REST API (/v1/memories, /v1/wake-summary, etc.) — Mem0-shaped, OpenAPI 3.1 spec at /v1/openapi.json
• Python SDK — from src.integrations.langchain_adapter import SCMClient
• JavaScript SDK — import { SCM } from "scm-memory" (Node 18+, Bun, browsers, Edge runtime)
• LangChain memory adapter — drop-in BaseChatMemory subclass
• Plain HTTP — POST /v1/memories from anything

Tool definitions exported in OpenAI / Anthropic / Gemini / OpenAPI formats from one source. See docs/INTEGRATIONS.md for 7 integration recipes.

---

What SCM is NOT

• ❌ Not a vector database. It uses one (NetworkX in-memory + SQLite/Postgres backing), but the value is the lifecycle, not the index.
• ❌ Not an LLM. Bring your own.
• ❌ Not a chat UI. It's the memory backend; the UI is your problem (or wire it to the included /static demo page).
• ❌ Not a fact-extraction prompt. Mem0 is essentially a prompt + a vector DB. SCM is a complete memory pipeline.
• ❌ Not 100% production-polished yet. v0.7.x is research-grade with strong tests; the deployment guide closes most rough edges. See docs/STATUS.md for an honest current-state read.

---

Honest comparison

We don't claim to dominate every memory benchmark. We dominate a different axis.

Capability	Vector DB	MemGPT	Mem0	A-Mem	SCM
Vector retrieval	✅	✅	✅	✅	✅
Working-memory bound	❌	❌	❌	❌	✅
Event-structured encoding	❌	❌	❌	❌	✅
Spreading-activation retrieval	❌	❌	❌	❌	✅
Contradiction-safe versioning	❌	❌	❌	❌	✅
Sleep-stage consolidation	❌	❌	❌	❌	✅
Schema extraction (REM)	❌	❌	❌	❌	✅
Wake-summary endpoint	❌	❌	❌	❌	✅
Curiosity-driven gap-filling	❌	❌	❌	❌	✅
Idle-aware autonomous learning	❌	❌	❌	❌	✅
Cleaner fact-recall benchmark (LoCoMo)	mid	mid	best	mid	weaker (we don't optimize for this)

LoCoMo, the standard agent-memory benchmark, rewards Mem0's architecture and penalizes SCM's. We document this honestly in the paper §9.5 and built LoCoMo++ (workload-sensitivity benchmark) showing the win pattern is architectural, not absolute.

---

Status

• 322 regression tests passing (pytest tests/ -q)
• 143 focused regression tests for Phase 7 + retrieval (pytest tests/test_*spreading* tests/test_*idle* tests/test_*curiosity* -q)
• 16/16 brutal LangChain harness scenarios passing (multi-day persona, contradiction, idle wake-summary, multi-user isolation, failure mode)
• 5,561× p50 latency speedup on add_memory since v0.7.2 (async ingest)
• 5-40× RAM saved at multi-user scale since v0.7.3 (embedding-model singleton)

See docs/BENCHMARKS.md for every measured number with reproduction instructions.

---

Architecture

                ┌───────────────────────────────┐
                │  YOUR AGENT / HARNESS         │
                │  (LangChain, Claude, custom)  │
                └─────────────┬─────────────────┘
                              │
                              ▼
        ┌───────────────────────────────────────────┐
        │  SCM CORE                                  │
        │  Wake-phase: encode → bind → retrieve      │
        │  Sleep-phase: consolidate → schema → gap   │
        │  Phase 7:  M1 idle daemon                  │
        │            M2 cross-session pool           │
        │            M3 schema extraction (REM)      │
        │            M4 wake-summary endpoint        │
        │            M5 curiosity engine             │
        │            M6 lifecycle policy             │
        └────────┬─────────────────────┬─────────────┘
                 │                     │
                 ▼                     ▼
       ┌───────────────────┐   ┌───────────────────┐
       │  LLM BACKEND      │   │  EMBEDDING        │
       │  Ollama/DeepSeek/ │   │  Ollama/OpenAI/   │
       │  OpenAI/etc.      │   │  sentence-trans.  │
       └───────────────────┘   └───────────────────┘

35-page paper at research/SCM_Final_Paper.pdf. Documents architecture, formal definitions for all 11 equations, brutal-testing methodology, ALB pilot, LoCoMo + LoCoMo++ honest comparisons, encoder-dependence analysis.

---

Project status

This project is in active development, working toward a public launch. See docs/STATUS.md for an unvarnished read of where it stands as a product, and docs/ROADMAP.md for what's coming next. The paper is ready and staged for arXiv submission but held until the product-readiness checklist completes — papers without products fade.

---

License

MIT. See LICENSE.

---

Contact

[e-mail]