| name | nowledge-mem |
| description | Use Nowledge Mem from Kimi Code for startup context, memory search, durable saves, thread search, and Kimi Code transcript import. |
Nowledge Mem is the user's cross-tool memory. Use it to start with the right context, recall prior work, save durable decisions, and make Kimi Code sessions searchable from other AI tools.
Startup Context
At the beginning of a meaningful session, or when resuming work, read Context Bundle through the CLI first:
nmem --json context --source-app kimi-code
This is intentional. Kimi Code's current remote MCP configuration accepts static HTTP headers but cannot map arbitrary process environment variables into headers. The CLI reads NMEM_AGENT_ID, NMEM_HOST_AGENT_ID, and NMEM_SPACE from the current Kimi process, so a named Raft/Kimi worker receives the correct AI Identity, Rules, default Space, and Working Memory.
Keep the non-empty authorship.agent_id, authorship.host_agent_id, and active_space.primary_space_id returned by Context Bundle for this session. Pass them as agent_id, host_agent_id, and space_id on later Nowledge Mem MCP calls. Never derive an identity from source_app.
If the CLI is unavailable but MCP is connected, call MCP read_context_bundle. If both are unavailable, use:
nmem --json wm read
Do not read both Context Bundle and Working Memory unless the user asks. Summarize only the parts relevant to the current task.
If nmem exists but rejects a Kimi Code command, flag, or MCP host helper, treat it as an outdated CLI rather than a broken Mem server. Check nmem --version, refresh the CLI from the same source, then retry. For the desktop-bundled CLI, ask the user to open Mem and run Settings -> Preferences -> Developer Tools -> Install bundled CLI. For standalone installs, use python3 -m pip install --user --upgrade nmem-cli or pipx upgrade nmem-cli.
Recall
Search memory when the user references prior work, asks for rationale, resumes a named project, investigates a regression, or asks about something that may already have been decided.
Prefer MCP when available:
memory_search for durable decisions, preferences, procedures, and learnings.thread_search when the user asks about prior conversations.thread_fetch_messages only after a thread result is relevant.
When startup Context Bundle returned an AI Identity or active Space, pass those
exact values on MCP search/read calls whose tool schema exposes the corresponding
agent_id, host_agent_id, or space_id argument.
CLI fallback:
nmem --json m search "what to look up"
nmem --json t search "conversation to find" --source kimi-code -n 5
For broad browsing across memories, threads, wiki pages, and artifacts, use the Knowledge Filesystem through MCP mem_fs when available, or:
nmem fs recall "topic" --in /memories -k 5
nmem fs grep "exact phrase" /threads
Save Durable Knowledge
When a meaningful decision, reusable procedure, user preference, correction, or non-obvious lesson appears, save it. Search first to avoid duplicates.
Prefer MCP:
memory_search for an existing memory.memory_update if the existing memory should evolve.memory_add for a new durable memory.
When startup Context Bundle returned an AI Identity or active Space, pass those
exact values on MCP writes whose tool schema exposes the corresponding identity
or Space argument. The server normalizes an explicit portable agent_id,
enrolls its profile once, and stores the attribution with a new Memory.
CLI fallback:
nmem --json m search "existing concept"
nmem --json m add "content" -t "Title" --unit-type decision -l "label" -s kimi-code -i 0.8
Use one strong memory instead of several weak notes.
Save Or Import Kimi Code Threads
Real thread sync is local to the machine where Kimi Code stores its session files. MCP is not the transcript-import layer.
If the user explicitly asks to save or import Kimi Code conversations, use:
nmem --json t sync --from kimi-code --session-id <session-id> --apply
To backfill older Kimi Code sessions, preview first:
nmem t sync --from kimi-code --limit 20
Then import:
nmem t sync --from kimi-code --apply
This works for local and remote Nowledge Mem because nmem reads local Kimi Code files and uploads normalized threads to the configured Mem server.
Status
When setup seems broken or the user asks whether Mem is connected:
nmem --json status
If nmem --json status works but Kimi-specific commands fail, do not keep using the old CLI. Upgrade it first, then rerun the failed command.
If the desktop app is on the same machine, nmem usually comes from the app. If Kimi Code runs on another machine, install the standalone CLI:
python3 -m pip install --user nmem-cli
Space And Identity
If the host process has NMEM_AGENT_ID, NMEM_HOST_AGENT_ID, or NMEM_SPACE, let nmem use those environment variables and retain the resolved Context Bundle values for MCP calls. Do not treat source_app=kimi-code as an AI Identity; it is only provenance.
User Overrides
For personal Kimi Code behavior, use Kimi's own AGENTS.md surface under $KIMI_CODE_HOME/AGENTS.md or the project instructions. Do not edit installed plugin files under $KIMI_CODE_HOME/plugins/managed/.
