Alpha MCPAPI reference · v1

The MCP your agents call.

Tritemporal memory that cannot lie about the past, a governed causal graph, knock-on simulation, and a decision ledger — the engine behind every Cascadian answer, at machine speed.

MCPPOSTGRESGOVERNED
Concepts

The Four Memory Networks

Cascadian's memory is four networks over one Postgres: temporal facts, a semantic layer of beliefs and lessons, an entity graph, and the decision ledger. Every row carries two time axes — when it was true, and when you could have known it — so the past can be reconstructed exactly as it was knowable.

FACTSBELIEFSCAUSAL EDGESPRIOR DECISIONSCONTEXT PACKtyped · pinnedAS-OF STAMPED≤2kTOKENS · L1+ CONTRADICTIONS · MISSING_EVIDENCE · ABSTENTION REPORT
Four networks → one typed context pack, every item as-of stampedILLUSTRATIVE

Two axes of time

as_of asks what was true; available_at asks what could we have known. The server stamps available_at itself on every write — backdating requires an audited reason. That one discipline turns "what did we know when we decided?" from a debate into a query, and makes lookahead bias detectable instead of deniable (invariant I2).

Append-only, by construction

Nothing is overwritten. A revision writes a new row, closes the old one, and links the two with a supersedes edge — in one transaction (invariant I1). Any past decision that leaned on the old belief still reconstructs exactly. The memory physically cannot lie about the past.

Reads that respect a token budget

bootstrap() returns a ~200-token session card. Search and packs return summaries plus ids; full cards come back only when asked for by id. No tool ever dumps raw chunks into your agent's context — L0, then L1, then L2.

Tools that use this