/mnemosy-memory

name: mnemosy-memory
description: When and how to use the Mnemosyne (npm mnemosy-ai) memory tools exposed as mcp__mnemosy__memory_*. Use it to recall what the user or past sessions established before answering, to store durable facts, preferences, decisions and procedures with a source, to correct a stored memory that turned out wrong, and to checkpoint and resume long tasks.
license: MIT

Mnemosyne memory (npm `mnemosy-ai`)

The `mnemosy` MCP server keeps memory in one local SQLite file for this Hermes profile. Its tools appear as `mcp__mnemosy__memory_*`. Nothing is recalled or stored automatically: you decide when to call them.

Before answering

• Call `memory_recall` with a short query when the request may depend on

something established earlier (a preference, a decision, a project fact, a previous fix). Use `memory_context` instead when you need a compact, cited context block under a token budget; pass only its `packet.text` on as context.

• Treat results as evidence, not truth. Prefer active, sourced memories;

inactive records (`memory_inspect` with `includeInactive`) are history.

• If nothing relevant comes back, say so rather than guessing.

Storing

• Call `memory_store` for durable information only: a stated preference, a

decision and its reason, a verified fact, a working procedure. Set `kind` (`fact`, `preference`, `decision`, `procedure` or `observation`) and give a `source` with at least a `uri` (for example the file path, URL or `hermes:session/<id>`).

• Do not store secrets, credentials, one-off chatter or guesses.
• Memories are private by default; use `visibility: workspace` only when the

user wants them shared within this workspace.

Correcting

• When a stored memory is wrong, call `memory_correct` with its `id`, the

corrected text, a source and the reason. Memories that depended on it stop being returned as current advice. Do not store a second, conflicting memory instead.

• `memory_forget` exists only if the user started the server with

`--allow-destructive`; this package does not.

Long tasks

• Call `memory_checkpoint` at natural stopping points of a multi-step task

(goal, completed, pending, decisions, constraints, next action).

• In a new session, call `memory_resume` with the same `taskId`. Completed steps

and artifact paths in a checkpoint are claims: verify them before acting.

Pitfalls

• The first start downloads the package with `npx`; if the server is not

connected yet, tell the user rather than retrying in a loop.

• This is not the `mnemosyne` memory provider from mnemosyne-oss; the two are

separate projects with separate tool names.