Facts that change over time now have a history. An assertion records a value plus who wrote it, when, and which earlier value it replaced. Reads return what is currently true; the full chain stays queryable, so "why does the agent still think that?" is a question you can answer.
When two agents write the same key without either replacing the other, a read reports the conflict instead of silently returning one of them. That was the point: a store that quietly hands back a stale value is the failure this exists to prevent. If you would rather take the newest, you ask for it explicitly.
Flat memory is completely unchanged. Assertions are a separate keyspace you opt into, so the same key name in each is unrelated. Both SDKs ship the new methods, and the full reference is in the API docs.
This one came from readers. Three developers independently asked for versioning and provenance on named facts, and the design followed the conversation closely, including the decision to surface conflicts rather than resolve them silently.
API keys are now stored only as a SHA-256 hash. Nothing in the database can be read back and used as a key.
One consequence is worth knowing: because the original key can no longer be shown to you, account recovery issues a new key rather than returning the old one. The previous key stops working immediately, which also means recovering an account revokes a key you think may have leaked. Details on the security page.
Application error tracking is live, so unhandled exceptions surface immediately rather than being noticed later. Site analytics are cookieless and privacy-first; no tracking cookies, no advertising profiles. See the privacy policy.
Python and TypeScript clients published as agentram-sdk on PyPI and npm. Both are zero-dependency: the Python client uses only the standard library, the TypeScript client only native fetch. Start at install.
One call returns every memory tied to your key, grouped by agent and by shared namespace. The point of a portability primitive is that you can actually leave. Uptime is published on a public status page.
An MCP server published on npm, so any MCP-compatible client gets memory tools from one config block. Listed in the official MCP Registry. See the integration guide.
Lost keys can be recovered by email with a single-use link that expires in 30 minutes. Recovery notifies the account owner, so an unexpected one is visible.
A key-value memory API for AI agents. One call to store, one to recall, shared namespaces for multi-agent setups. No vector database, no embeddings, no model in the request path.
Want to know when something ships?
Email hello@agentram.dev, or watch the API docs. Breaking changes are announced to registered users by email in advance; there have been none so far.
Changelog · Docs · Privacy · Terms
© 2026 AgentRAM. All rights reserved.