Search memory store when past insights would improve response. Recognize when user
Run one targeted search before concluding continuation, review, regression, release, connector, prior-decision, cross-tool, or exact-history work. References to a previous fix, "that approach", or "like before" are strong signals. The startup Context Bundle / Working Memory is a briefing, not a substitute for this search. Do not wait for the user to explicitly request retrieval.
Skip trivial, independent tasks, generic syntax questions, and an explicit fresh start.
Before searching, briefly tell the user what prior context you are looking for. Use the CLI; this plugin does not register an MCP server. If NMEM_CLI_PATH is set, use its quoted value as the executable instead of bare nmem.
Start with the smallest relevant surface:
# Durable decisions, procedures, and learnings
nmem --json m search "3-7 core concepts" -n 5
# Prior conversations and exact discussion history
nmem --json t search "query" --limit 5
Keep the active Space from the Context Bundle. NMEM_SPACE selects an explicit session-wide lane; legacy NMEM_SPACE_ID is compatibility-only. If an AI Identity resolved another lane and no session override is set, pass that existing lane with --space "<space name>". Never infer a Space from cwd, git, or a project name, and never remove a selected lane because it returned no matches.
If normal memory search is weak or conceptual, try Deep Search once:
nmem --json m search "query" --mode deep -n 5
Inspect only relevant hits. Judge their content and server warnings rather than applying one fixed score threshold across search modes. When a memory includes source_thread, or thread search finds the likely conversation, inspect one small targeted range for the specific messages needed:
nmem --json m get <memory_id>
nmem --json t show <thread_id> --limit 8 --offset 0 --content-limit 1200
Use a known target --offset once; do not advance through pages.
Use label, importance, or time filters only when they narrow this question. For code-backed knowledge,
verify the finding against the current repository before treating it as current.
Use retrieved knowledge naturally and cite the memory or original thread when it supports the answer. Distinguish a successful search with no relevant matches from a CLI, authentication, timeout, or server failure. If retrieval is unavailable, say so briefly and continue the authorized work.
A session-start receipt confirms briefing loading. A Bash search or Skill call shows actual retrieval in the host transcript; do not claim a search occurred because startup context was loaded.
Run /nowledge-mem:status or nmem status to check the connection. If the CLI is outside PATH, set NMEM_CLI_PATH to its executable path. Install nmem-cli with pip or pipx if needed; Arch Linux users can use yay -S nmem-cli or paru -S nmem-cli. For remote Mem, configure the shared client on this machine with nmem config client set url and nmem config client set api-key.
Search memories or threads first, then choose the specific conversation and message range needed. Call thread_fetch_messages, a host thread-fetch tool, or nmem --json t show at most once per question, using a small message limit (for example 8) and a content limit of 1200 where supported. A known target range may use a nonzero offset once.
Never loop over offsets to reconstruct a conversation. Do not raise --content-limit to compensate for a slow or incomplete read. Small output limits do not guarantee cheap server-side work on every deployed backend. If a read is slow, incomplete, fails or times out, stop, report the evidence gap, and refine the search rather than retrying pages. A later explicit user request for another range is a new bounded read; a routine “continue” is not permission to drain the thread.
October 25, 2025