scm-memory 0.9.1

SCM: local-first lifecycle memory for AI agents — wake, sleep, consolidation, and recall.

SCM: Sleep-Consolidated Memory

SCM is plug-and-play lifecycle memory for AI agents.

Most memory layers store facts and retrieve them later. SCM keeps that simple surface, then adds the human-like part that normal retrieval misses: memory can sleep, consolidate, forget stale noise, preserve current facts through contradictions, discover schemas, and wake with a summary of what changed.

Try It In 60 Seconds

python -m venv .venv
source .venv/bin/activate
pip install scm-memory
scm demo

Open the printed local URL, add a few facts, run sleep, and read the wake summary. No OpenAI key, hosted account, or vector database is required for the first run.

Check your install:

scm doctor

Use SCM In Your App

from scm import SCMEngine

memory = SCMEngine(session_id="demo", sandbox=True, offline=True)

memory.add_memory("The user is building a lifecycle-memory product.")
memory.add_memory("The product must stay local-first and plug-and-play.")
memory.sleep("deep")

result = memory.search_memory("what is the user building?")
print(result["memories"][0]["description"])

print(memory.wake_summary()["narrative"])

Universal Five-Tool Contract

SCM exposes the same contract through Python, CLI, REST, MCP, JavaScript, and LangChain/LangGraph:

Tool	Purpose
add_memory	Store a fact, event, preference, or observation.
search_memory	Retrieve memories by associative recall.
sleep	Run consolidation, schema extraction, forgetting, contradiction handling, and curiosity.
wake_summary	Report what changed during idle/sleep time.
forget	Suppress a specific memory by id.

consolidate remains a backward-compatible alias for sleep.

Product Surfaces

scm demo
scm chat
scm serve --host 127.0.0.1 --port 8000
scm mcp
scm sleep --mode deep
scm wake-summary --hours 24

Python:

from scm import SCMClient, SCMEngine

REST:

curl -s http://127.0.0.1:8000/v1/tools?format=openai

JavaScript:

npm install scm-memory

import { SCM } from "scm-memory";

const scm = new SCM({ baseUrl: "http://localhost:8000/v1", userId: "demo" });
await scm.addMemory("The user likes concise technical answers.");
const recall = await scm.searchMemory("communication preference");
console.log(recall.memories);

MCP:

scm mcp

Use it with Claude Desktop, Cursor, VS Code, or any MCP-compatible agent.

What SCM Is Not

SCM is not a vector database wrapper, not a chatbot UI, and not a claim that sleep-style memory wins every retrieval task. Retrieval systems remain strong for fixed snapshot lookup. SCM targets lifecycle workloads: interference, contradiction, schema discovery, curiosity gaps, and wake summaries after idle time.

Docs

• Getting Started
• Product Thesis
• Architecture
• Integrations
• Reproducibility
• Artifacts

Research

The paper is proof for the product, not the first-use path.

• Paper: research/SCM_Final_Paper.pdf
• Reproducibility: docs/REPRODUCIBILITY.md
• Release artifacts: GitHub Releases

Citation

@techreport{shinde2026scm,
  title = {SCM: Lifecycle Memory for Language Agents},
  author = {Saish Shinde},
  institution = {Clyrai IP Studio},
  year = {2026},
  month = {July},
  url = {https://github.com/clyrai/SCM_OpenSource}
}

License

MIT