When an agent needs to remember.
SyntheticMem is a network memory any MCP client can mount. A few shapes it fits — each one just a namespace and a key away.
Cross-session dev memory
A coding agent learns a gotcha mid-task — "staging needs PG_SSLMODE=require" — and writes it with a source and an invalidation condition. The next session, or next week, recalls it by meaning instead of re-deriving it. The fix outlives the window it was found in.
Shared memory across agents
Several agents — a planner, a builder, a reviewer, across Claude Code, opencode and grok — point at the same endpoint and namespace. What one agent learns, the others can recall. One remote store replaces a pile of per-agent scratch files.
Provenance-audited ops knowledge
Operational lessons ("retry the deploy if the pooler refuses TLS") carry their source and the condition that would make them wrong. When the platform changes, the stale rule is superseded, not silently overwritten — the audit trail of what was once believed survives.
A 24h scratch memory
A one-off or throwaway agent mints an anonymous demo key and gets a working memory in 30 seconds — no account, no cleanup. Everything it writes auto-expires in 24 hours. Register later to promote the same workflow to permanent.
Born from the pixel-office thesis
SyntheticMem re-homes a thesis already proven inside pixel-office — that agent memory can be persistent, shared and recallable across sessions and across agents, with provenance, invariants and supersede-not-delete. There that memory was coupled to one app. Here it is a pure MCP service: no office app, no CLI, no local workspace files — memory lives in a remote database and recall is a tool call, reachable by any MCP client alike.