v0.13.0

Added

• memory_graph, an eleventh tool: reads the association edges around given
  concepts. Wraps the live POST /memory/graph (core's
  GraphRequestSchema/GraphResponseSchema). Input: seeds (1-20 concept
  names, each ≤200 chars, non-blank), depth (1-3, default 1), domain
  (an xroom: address reads that room's own association graph instead of
  yours — the ONE mechanism it shares with memory_read's domain; any
  other value is inert here, since hebbian_edges has no domain column),
  min_weight (≥0, honoured at every hop with no server override, 0
  included) and limit (1-500, default 100, mirrors memory_read's
  top_k bounds). Output: nodes (concept, degree — edges in THIS
  response touching it, not the concept's whole-store degree), edges
  (source, target, weight, valence, count, updated_at),
  truncated, and min_weight_applied (the floor actually used — the
  engine floors at 0.05 from every hop of a depth ≥2 walk when the caller
  left min_weight unset). Text rendering sorts edges by weight,
  strongest first, and states truncated/min_weight_applied as a
  trailing note when either is non-default; an unreadable 2xx (any of the
  four required response fields absent or mis-shaped, on any node or edge)
  answers isError rather than rendering a partial or fabricated graph,
  the same rule every other tool here applies to its own response shape.
  Annotations: readOnlyHint: true, destructiveHint: false,
  openWorldHint: false. The server instructions in src/teaching.ts now
  name it too (memory_graph shows links); fitting that mention under the
  800-char cap shortened two existing sentences there — see the comment
  above SERVER_INSTRUCTIONS. Every sentence that states the tool count
  (the README's stability-contract line, llms-install.md,
  docs/shared.md, the src/tools.ts headers) says eleven in this
  release; the tools.json entry below says how they are held to it.
• tools.json, the tool list as a generated, published artifact.
  scripts/generate-tools-manifest.mjs registers the tools from the built
  server (dist/shared.js) on an in-memory McpServer, reads tools/list
  through a real client, and writes tools.json at the package root: name,
  title, description and annotations per tool, sorted by name, with the
  package version and the count. It is generated by npm run build
  (postbuild), shipped in the npm tarball (files) and reachable as
  @mnemoverse/mcp-memory-server/tools.json (exports), so a downstream
  copy of the surface can be derived from it or checked against it instead
  of typed by hand. The .mcpb manifest.json tools[] is now generated
  from this file (name and description, the MCPB shape) rather than from a
  hand list in src/configs/source.json, which had ten entries while the
  server registered eleven; the hand list is gone. A stale tools.json is
  caught by test/tools-manifest.test.ts on every PR and by the release
  workflow, which compares the committed file with the one the build
  regenerates (verify:configs only checks the other artifacts against it).
  test/tools-manifest.test.ts pins the committed file to a
  live in-memory listing, pins manifest.json to it, and checks that every
  sentence stating the tool count (README, llms-install.md,
  docs/shared.md, the src/tools.ts headers) states the count
  tools.json states, eleven in this release.

Removed

• memory_feedback's atom_ids parameter, as announced. It was renamed
  to memory_ids in 0.11, kept as a deprecated alias for one full MINOR, and
  its removal moved from 0.12 to 0.13 (see 0.12.0). atom_ids is no longer
  a parameter: a call that passes it instead of memory_ids is refused as
  invalid arguments naming memory_ids, and nothing is sent; passed next to
  memory_ids, it is ignored like any unknown field on every tool here, and
  the ids come from memory_ids alone. With one name left, memory_ids is
  schema-required, and the handler's two refusals for "both names" and "no
  ids" are gone with it.
  The request to the engine is unchanged: it still carries the list as
  atom_ids, the engine's own field name.

• npm: https://www.npmjs.com/package/@mnemoverse/mcp-memory-server

• MCP Registry: https://registry.modelcontextprotocol.io/v0.1/servers?search=mnemoverse

What's Changed

• feat(tools): memory_graph reads the association edges around given concepts by @izgorodin in #174
• ci(release): start consumers' contours with a workflow dispatch (mnemoverse/ai-sdk) — for Σ review by @izgorodin in #177
• feat(configs): tools.json, the generated tool manifest the .mcpb and every count derive from by @izgorodin in #179
• ci(release): the wave tracker, a Wave vX.Y.Z issue per release ticked daily from live probes by @izgorodin in #178
• feat(feedback)!: remove memory_feedback's atom_ids, as announced for 0.13 by @izgorodin in #183
• release: 0.13.0 by @izgorodin in #182

Full Changelog: v0.12.1...v0.13.0