diff --git a/README.md b/README.md index 2babe55..bc95dc6 100644 --- a/README.md +++ b/README.md @@ -117,6 +117,54 @@ Until latency benchmarks are linked from the docs, treat the engine as "designed for single-digit-ms local retrieval on a developer laptop at typical agent corpus sizes" — a design target, not a guarantee. +## Validation boundary + +Different kinds of proof live in different places. This map states what this +public repository's CI proves on its own and what is validated elsewhere, so +readers do not assume the public repo independently proves every product claim. + +**Proven in this repository's public CI (every pull request):** + +- repo hygiene +- package metadata checks +- affected build, typecheck, lint, and self-contained package tests (Node 22 + and 24) +- code-health gates +- package `pack` dry-run plus tarball-shape verification +- docs contract (install commands, package status labels, and smoke rows stay + in sync) +- public integration smoke +- security compliance + +**Also in this repository, run in package or release contexts (not the per-PR +affected lane):** + +- Core OpenAPI generation and drift check (`generate:openapi` / `check:openapi`) +- Core API schema tests (Schemathesis) +- Core Docker image smoke (runs in the Docker image publish workflow) +- DB-backed Core tests, which require Postgres/pgvector provisioning + +**Validated outside this repository:** + +- hosted-service behavior +- release orchestration and marketplace operations +- full benchmark and evaluation harnesses, including performance and cost + reproduction +- global, end-to-end product behavior + +These run in separate internal repositories and surface as public artifacts when +they are ready to publish. + +**Not yet public (planned):** + +- reproducible benchmark artifacts and raw evaluation harnesses, including the + BEAM and LoCoMo suites referenced in the headline results +- any global lifecycle fixtures not currently published + +This is a map, not a ranking: external or not-yet-public validation is not +presented as equivalent to public CI from a reader-trust standpoint. It simply +lives in a different place. + ## Quickstart For the full walkthrough, see the diff --git a/packages/mcp-server/README.md b/packages/mcp-server/README.md index c5d2e84..10bc408 100644 --- a/packages/mcp-server/README.md +++ b/packages/mcp-server/README.md @@ -7,6 +7,20 @@ MCP server that exposes [AtomicMemory core](../../packages/core) as four tools t - `memory_package` — token-budgeted context package - `memory_list` — list recent scoped memories +## Authoritative contract + +The REST API and the [`@atomicmemory/sdk`](../../packages/sdk) type surface are +the authoritative memory contract — provenance, scope, mutation decisions, +lineage, retrieval scores, and context-package metadata are defined there. This +MCP server is a thin callable-tool adapter over that contract. + +Tool results are returned as JSON-stringified text for host compatibility, so +the text payload is a transport convenience, not a separate audit surface. For +evidence or audit purposes — inspecting why a memory exists, what it replaced, +or what shaped a retrieval — read the REST/SDK projection rather than parsing +MCP tool text. New memory semantics land in Core and the SDK first; this adapter +exposes them, it does not define them. + ## Status: package entrypoint This package is intended to publish as `@atomicmemory/mcp-server`. Cursor and