The first agent memory layer with a wake phase and a sleep phase. Your agent doesn't just retrieve facts — it consolidates, abstracts patterns, and surfaces what it figured out while you were idle.
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.
Every other agent memory product — vector stores, RAG layers, context windows — only models the wake phase. They store facts and retrieve facts. They never think between sessions.
SCM models both phases. Once per night, in the user's configured timezone, your agent runs a real consolidation cycle: schemas form, contradictions resolve, low-value memories fade. The next time the user comes back, the agent has grown.
While the user is active, SCM behaves like prefrontal working memory: bounded, cue-driven, contradiction-aware.
At the user's bedtime, in their timezone, SCM runs NREM consolidation and REM schema extraction. Once per night.
Six modules, one HTTP API. Every layer is exposed and configurable — but you don't have to touch any of them to get the default product behavior.
An attention gate filters incoming text — "hmm okay" gets dropped; "I'm allergic to peanuts" gets encoded with high importance. Concepts go in, noise stays out.
Cue-driven graph propagation, not raw vector similarity. Asks "where do I work?" returns "I'm at Atlas Labs" even though those tokens never overlap.
"I left Northstar, I'm at Atlas now" doesn't delete the old fact. It supersedes it. Ask "where did I used to work?" and the agent still knows.
NREM strengthens rehearsed memories and lets weak ones fade. REM finds patterns: "Tuesday-morning runs are recurring." Both phases run during the user's sleep window.
The next time the user shows up, your agent surfaces a short narrative of what it consolidated and what patterns it noticed. The most-shared SCM moment.
Every user_id is namespaced under your account server-side. Cross-account reads are impossible — even if a customer guesses another customer's user_id.
One Python SDK. One HTTP API. One MCP server. Works with LangChain, OpenAI Agents, Claude Desktop, Cursor, or anything that can issue an HTTP call.
Each request goes to your account's /v1/memories endpoint with a Bearer key. Rate limits are per-key. Tenancy is enforced server-side.
Anywhere your AI is supposed to know more about the user this week than it did last week.
Remembers what the user said weeks ago. Surfaces patterns ("you've mentioned this three times now"). Wakes up with a summary of what it figured out.
One ticket today, another in three weeks — same customer, same context. SCM consolidates the relationship, not just the conversation.
Knows what you're building, what you tried last Tuesday, what you decided not to do. Doesn't suggest the same dead-end twice.
The full architecture, the brutal-test methodology, the workload-sensitivity benchmark, and the four real bugs we caught in the process. Honestly written, including the failure modes.
Read the paperYou bring your own LLM key. We never proxy or charge for inference. We charge for the memory infrastructure — eventually. Right now, all of it is free.
Sign up, mint an API key in 30 seconds, drop it into your LangChain code. Your agent learns the way humans do.