Frequently asked questions.

SyntheticMem is a pure MCP server that gives any AI agent permanent, shared memory with semantic recall. The difference from naive RAG-memory is the trust model. RAG-memory stores text, embeds it, returns the top-k matches, and trusts them. SyntheticMem keeps the embedding index but refuses the blind trust: every memory is a row that carries its source, confidence, status and the condition that would make it wrong, and recall returns that provenance alongside the content so the agent can verify instead of assuming. Embeddings are the index, not the truth.

Any MCP client that speaks JSON-RPC 2.0 over Streamable-HTTP. That includes Claude Code, Claude Desktop, opencode and grok, as well as a raw HTTP client with no SDK at all. There is also a REST v1 API that mirrors the same tools for plain HTTP use.

Point your client at https://syntheticmem.com/api/mcp with a per-agent agt_ bearer key. In Claude Code that is one command — claude mcp add --transport http — or a .mcp.json file in your project root. opencode uses a "type":"remote" block in opencode.json, grok uses grok mcp add or the same .mcp.json, and a raw client just sends initialize then tools/call with the Authorization header. The Connect section on the home page has copy-paste snippets for each, and the demo-key dialog injects your key straight into them.

Yes. SyntheticMem is free while it is in preview. You can start anonymously with a 24-hour demo key and no sign-up, or register a free account for permanent memory. There is no paid tier yet, and no card is required.

An anonymous demo key is not linked to any account, so the memories it writes are capped and automatically expire after 24 hours — it is meant for trying the server or for throwaway agents. When you register and link an agent to your account, that agent's memories are permanent, unless you give a memory an explicit time-to-live or delete it yourself.

Yes. Memory lives in a remote PostgreSQL database, not in a context window or a local file, and it is scoped by agent and namespace. A fix learned in one session is recallable from the next, and several agents or clients that share the same key and namespace can recall what any of them wrote. This cross-session, cross-agent recall is the whole point.

Embeddings are computed locally on the server with Ollama running the nomic-embed-text model. Your memory text is embedded on the SyntheticMem server itself and is not sent to any third-party embedding API. The vectors are stored in PostgreSQL with pgvector.

Provenance is the metadata that travels with every memory: its source, when it was learned, when it was last observed, how many times it was compressed (hop_count) and the namespace it held against. Invalidation is falsifiability — an optional tag and a concrete condition that would retire the memory, for example "retire if the platform ships a retry path." Confidence is a numeric score that is preserved through recall and consolidation and is never silently inflated.

Every memory is a fact, a hypothesis or retired. A sourced save becomes a fact; an unsourced save is demoted to a hypothesis ("do not execute from this"). A hypothesis is promoted back to a fact only by a fresh observation (memory_observe), never because it has been used a certain number of times. Superseding or forgetting a memory retires it — the old row is kept and addressable but hidden from default recall. Nothing is hard-deleted, so the audit trail of what was once believed survives.

Agents are isolated from each other by per-agent keys and a fail-closed scope, so one agent cannot read another agent's memories, and namespaces sub-scope memory within an agent. Keys are stored only as salted hashes. The website uses Google Analytics and Microsoft Clarity for aggregate usage analytics, which set cookies — see the Privacy page for the full disclosure and how to opt out at the browser level.

Registered users manage and revoke per-agent keys from the dashboard at /app. A revoked key stops working immediately for every tool and REST call. Anonymous demo keys cannot be revoked individually because they are not tied to an account, but they are rate-limited and their memories expire within 24 hours.

Consolidation summarizes many episodic memories into a single semantic memory while deliberately preserving the doubt. SyntheticMem clusters episodic rows by namespace and writes one consolidated row whose confidence is the minimum of its sources (never inflated), whose invalidation conditions are a deduped union of the originals, and whose status is hypothesis because a derived summary is not a fact. The source rows are superseded, not deleted. A summary that loses the "maybe" is a eulogy, not a record.