v0.14.0

The server now says what each tool does and when it applies, and no longer
tells the model how to behave: the form Anthropic's connector directory asks
for. The rules that used to live in the tool text come from a new prompt,
setup_memory, into the user's own agent instructions.

Added

• setup_memory prompt. It renders memory rules for the user to place
  in their own agent's instructions: CLAUDE.md, AGENTS.md, a Cursor rule, or
  Instructions for Claude in Claude's settings. An optional host
  (claude-code, claude-ai, cursor, codex) narrows it to one place.
  Like the other prompts, it calls nothing.

Changed

• The server says what each tool does, not how the model should behave.
  The server instructions and eight tool descriptions no longer tell the model
  to act "proactively", to check "ALWAYS", "as a habit", or "don't wait to be
  asked", and no description tells it which tool to call next. Anthropic's
  connector directory rejects descriptions that "tell Claude how to behave",
  and its submission form asks the publisher to confirm that tool descriptions
  "contain no instructions about model behavior, other tools, or external
  instruction sources". The descriptions still say when each tool applies.
  The rules that used to live here now come from setup_memory and the host
  plugins, where the user places them. A test bans the removed words on every
  surface the server advertises.

• memory_feedback and memory_read name the moment a rating applies:
  after an answer that relied on or rejected recalled memories, with the ids
  memory_read returned. Measured in Claude Code before release: with the pushes removed and
  no such sentence, ratings fell from 17 of 30 answers to 6; with it, 22 (18 in a
  recheck of the final wording, which also covers rejected memories). Tests pin
  both sentences.

• memory_graph states in its description what its parameters already
  said: domain has an effect only for a room's address, and at depth 2 or 3
  edges below weight 0.05 are dropped unless min_weight is set.

• memory_stats no longer leads with "episodes vs consolidated
  prototypes": consolidation is not running on the hosted service. The output
  fields are unchanged.

• Output field descriptions say what is true and stable (#191). The
  importance score of memory_write, and its "NOT STORED" reply, call it an
  approximate hint that can read lower for non-English text; the
  language figures they quoted were measured on an embedder the service no
  longer uses. memory_stats describes episodes, prototypes and
  hebbian_edges as what a user can check today, and its reply says
  "N learned concept links". memory_feedback's coactivation_edges is
  worded plainly. Field names are unchanged.

• The empty-store greeting states instead of instructing (#191): it says
  what the store is and how a memory gets into it, with no "save the first
  memory now". Directory reviews treat instructions inside a tool result like
  instructions in a description.

• 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: say what each tool does, not how the model should behave; add setup_memory by @izgorodin in #190
• fix: output field descriptions and the empty-store greeting say what is true and stable by @izgorodin in #192
• release: 0.14.0 by @izgorodin in #193

Full Changelog: v0.13.1...v0.14.0