Memory as a network service.
SyntheticMem gives any AI agent permanent, provenance-aware memory and semantic recall — as a pure MCP server, reachable by any client over one endpoint.
Where it came from
The thesis — that agent memory can be persistent, shared and recallable across sessions and across agents, with provenance, falsifiability and supersede-not-delete — was first proven inside pixel-office, where memory lived as a feature coupled to one app: its auth, its presence layer, its CLI bridge.
SyntheticMem re-homes that same thesis as a pure MCP service. There is no office app to run, no CLI to install and no local workspace files. Memory lives in a remote database and recall is a tool call, reachable by Claude Code, Claude Desktop, opencode, grok or a raw HTTP client alike. The idea — "the agent forgets, the memory doesn't" — stays; the coupling goes.
What it is, concretely
- A Laravel app on syntheticmem.com exposing an MCP server at POST /api/mcp and a REST v1 mirror.
- PostgreSQL 16 with pgvector for native vector search, and local Ollama nomic-embed-text embeddings so memory text stays on the server.
- A provenance- and falsifiability-first memory model: a fact/hypothesis/retired status machine, supersede-not-delete, confidence preserved through doubt-preserving consolidation.
- Per-agent keys, namespaces, a self-serve 24-hour demo, and a dashboard at /app for permanent memory.
Who builds it
SyntheticMem is built by autoaidev. The source lives on GitHub at github.com/autoaidev/syntheticmem-app, part of the autoaidev organization. It is in preview and free to use — see pricing.