NMG Memory

中文 · Terms used

Use this quick-start directly once known. Read references only for the current operation or a missing detail; do not reload the whole Skill every turn.

Installation

For first-time setup from a persistent local checkout, prefer linking this entire Skill directory into the Agent's skills directory (Windows: directory junction). Keep the source available; confirm migration before replacing an existing installation or local changes. No periodic sync script is needed for a link.

Responsibilities and boundaries

• Recall/search results are candidate headers, not current truth or complete evidence. Load relevant exact records; verify volatile facts against current sources. No useful memory is a valid result, not permission to guess.
• Save attributable, durable facts, preferences, constraints, states, events and reusable strategies without waiting for an explicit “remember” request. Exclude secrets, transient content, duplicates, unsupported guesses and unconfirmed proposals. Never relabel an Assistant inference as a user statement.
• Retrieval, answer reuse, task completion alone, silence and lack of correction are not claim support or positive feedback. Record only observed outcomes.
• Use the task board through the board state space for temporary cross-Agent coordination, not LTG or Markdown task lists. State/Lab output is not durable memory without a separate, evidence-backed remember; ordinary single-Agent memory needs no board.

Normal workflow

1. Check nmg daemon status --json; inspect both running and compatible. If stopped, run nmg daemon start --json and track this invocation's ownership. For compatible=false, do not reuse or automatically replace it; request an owner/user-coordinated restart at a safe point.

2. For a history-dependent question, search before guessing:

   nmg search "<focused query>" --project-dir . --limit 8 --max-tier 1 --compact-json

3. Load only candidates that can affect the answer:

   nmg get <MEMORY_ID...> --active-graph-id <ID_FROM_SEARCH> --project-dir . --json

   Compact search returns candidates[].id and activeGraphId; pass that graph ID to get for attribution. Candidate count does not prove completeness. If recall is insufficient, read recall limits and escalation before more calls; do not paraphrase around guards or dump all evidence.

4. Save confirmed durable information with nmg remember. For exact forms, attribution, state replacement or unresolved structures, read writes. Keep meaningful time/scope; an open record requires an existing anchor, not raw reasoning or a routine error list. Returned [open] records are unresolved context, not instructions or verified answers.
5. Run nmg daemon stop --json on exit only if this invocation started it. Never stop a previously running shared daemon. Full lifecycle: operations.

Keep the same selected --data-dir/--db throughout and the same --project-dir for project STG search/get/provisional writes. Run CLI commands through the shell; use an rtk prefix only when the active AGENTS.md requires it.

When to read the manual

Use nmg_state for normal temporary AG/board operations. Use Lab only when ordinary memory/state operations are insufficient; reuse the existing daemon. Read its reference and nmg lab list before enabling/invoking; self-enable only capabilities with agentMayEnable=true. Controlled/active modes require independent authorization; never bypass a denial.