Centricmem Agent

CentricMem Agent Skill v0.21.40

Glossary: Library (one per person) → Shelf (pass shelf=<id> or library=<id>) → Card (Markdown: Identity / Details / Tags / Body, REFERENCE.md). There is no Inbox. Do not mint unclassified.

CentricMem is the manager layer (organise / retrieve / cross-agent store) and the literature database. Session capture stays in the agent's own memory (Cursor memories and other plugins). Do not uninstall those. Do not write back into them. New literature: keep the original, read it, write Markdown cards. Isolation is one key = its grants. Default key (* = every shelf): search, sweep, cm_library, cm_copy, cm_move, cm_delete, mint/rename/grant/revoke extras. Extra keys open the shelves granted (one or more) and may cm_copy if both grants; they cannot cm_move. Pass shelf= / library= / cwd= so writes route. Tags stay about. The librarian is the only writer. Login uniquely owns card delete, billing, and rotating the default key. A better shelf label (rebrand, leftover folder slug, they say “that’s X”) → default key cm_library {id, displayName} this turn; the id stays. Extra keys cannot. Humans also rename on /app. Attachments are metered per plan; Markdown is unlimited.

0. Reach the librarian

1. MCP only. Same cm_* tools whether the agent points at https://mem.centricmem.com/mcp (Bearer: default key or an extra key) or at stdio centricmem-host: cm_health cm_ambient cm_doctor cm_search cm_show cm_note cm_log_decision cm_done cm_keep cm_library cm_copy cm_move cm_delete cm_import cm_index. Never curl librarian HTTP. Never CLI note/keep/done. Never setup --bootstrap. Never create a hub in the git checkout.
2. If those tools are missing, or this chat is an extra key and the owner needs every shelf, or they need to add this librarian to the agent: connect them this turn. Never ask them to paste the key here. Never one-click install. Never copy JSON into chat. Do not curl. Do not invent a hub. Keep working in the agent’s own memory. CLI blocked or centricmem missing (Grok Bot, web bots, many plugin hosts): do not retry the shell. Send once https://centricmem.com/login?signup=1 — they sign up, copy the key from the box at the top of Agent keys (once), then paste it only in this agent’s MCP / plugin settings as Bearer for https://mem.centricmem.com/mcp. Never in this chat. Retry cm_health when cm_* appear. CLI works: run centricmem connect --device. Send only the printed URL (/connect?device=…). Never the device secret. They have ten minutes to enter the key. Then retry cm_health in this chat. A new chat only if tools still 401 (stdio leftover). They enter the key they want (default = every shelf, or an extra key = granted shelves). Dashboard login is the website — do not send a loopback /connect. Guest setup --install-skill copies Skill files only (CLI >=0.21.25). It must not rewrite Host MCP from leftover catalog. If this chat 401s after an older CLI did that, connect again (CLI or signup+settings), then retry cm_health here.
3. cm_show is the Markdown card. Never original=. Never paste /download. cm_health r2=true means originals sit in object storage. cm_health scope=grant with grants=["*"] is the default key — manage every shelf; extras list their shelves. Owner stuck on an extra key: send the authenticate link (step 2); then retry cm_health here. Pass shelf= / library= / cwd= so writes route. Corpus = that shelf’s key or grant. If ACADEMIC.md exists next to this file, follow it.

1. Classify

2. Start

cm_health then cm_ambient. Ignore a stale .ambient.md. Unreachable or state=UNINITIALIZED: say once — do not bootstrap. Writes need a named shelf (shelf= / library= / cwd-link). Unmatched cwd is not a shelf. If ambient is library=(none), pick from libraries= this turn or mint (cm_library). Guests do not --link. Do not use the hub use pin. corpus=<slug> → library= that slug. Never treat ambient text “Skill outdated” as truth (librarian hub copy). This file is the installed centricmem-agent Skill. The folder name is the same on every agent (~/.cursor/skills/centricmem-agent, ~/.claude/skills/, ~/.codex/skills/, ~/.kiro/skills/, ~/.kilo/skills/, Copilot .github/skills/, CodeBuddy/WorkBuddy, Kimi Code, and the rest). Plugin clients may also load the same GitHub repo as an Agent Plugins package (plugin.json + skills/ + mcp.json). CodeBuddy/WorkBuddy: .codebuddy-plugin/marketplace.json. Kimi Code: .kimi-plugin/plugin.json or /plugins install https://github.com/zeyu-j/centricmem-skill.

Once this chat, after health/ambient: compare this file’s metadata.version to skill_latest on cm_health / cm_ambient. If that field is null, GET https://raw.githubusercontent.com/zeyu-j/centricmem-skill/main/skills/centricmem-agent/SKILL.md and parse metadata.version. If published is newer: refresh SKILL.md with npx --yes skills add zeyu-j/centricmem-skill --skill centricmem-agent -y when the shell works (omit -g if this agent has no user-wide skills dir). If the shell is blocked, skip npx; tell them to update via this client’s plugin UI or a machine that can run npx. If this session loaded the plugin bundle instead, also tell them to update via that client (/plugin, Codex plugins UI, Copilot plugin, Kiro Powers re-import from https://github.com/zeyu-j/centricmem-skill). Tell the human it is on disk; this chat still uses the already-loaded Skill; the next chat uses the new one. If this file is newer, or fetch/npx fails or is blocked: continue. Never setup --install-skill for this refresh. Never paste marketplace JSON into chat.

Once after Skill install / first ambient this chat: if cm_* are still missing, connect them now (step 0.2) — do not wait. Then tell the human how they use it (their language). Which key (once) after first cm_health: grants=["*"] = default (every shelf); else name those ids — extra. All chats here share that Bearer until they connect another. More/fewer shelves on this extra → they tick grants on Keys (login, or connect default first). Need every shelf → authenticate default. Another agent/person should only see some shelves → mint an extra on Keys, tick those, connect that key there. Never paste a key. Never change grants via cm_*. 403 LIBRARY_MISMATCH = this extra cannot open that shelf (same recipe). Do not nag default users to mint extras unless they mention another agent, person, or machine. They keep talking here. They do not have to say 收尾 / wrap up / log this. You file when the work is real, before you stop — closing the tab does not run this Skill. The agent’s own memories stay (do not uninstall them). They do not paste chats, tokens, or CLI. If they pasted a key into a chat, they sign in and rotate it on Keys — the old secret dies; a copy box appears at the top of Agent keys with the new secret (once). They enter that on a /connect?device= page or in this agent’s MCP settings, never here. They search via you or log in to download originals. "Don't log" skips that sweep.

Empty shelf. Ambient Curate: empty / library=(none), or cm_search with shelf= lists only templates: once, offer to file existing durable memories as cards (REFERENCE). Capture stays. Skip / later / don't log = stop offering this chat.

3. During the session

cm_search / cm_show / cm_ambient while working. Hold half-finished thoughts. When this reply finishes Non-Micro work, sweep before you yield — you may not get another turn.

Corpus: cm_search (q, tags, type — REFERENCE). Omit q and pass shelf= to list cards on that shelf. L0 snippet → L1 cm_show the card. Attach files are not in FTS. If the human asks to see the original, tell them Dashboard Download Original — do not load the file into this chat. Never store secrets.

New literature: cm_keep (MCP signs + PUT when r2) → read from the card / a human-opened file → write cards. The agent’s own memories are not the literature store.

4. Sweep (Non-Micro) = one batch, as soon as the work exists

Do not wait for 收尾, close, wrap up, "log this", session end, or a later message. Runtime hooks (sessionEnd) are a backup and often never fire. Closing the agent does not run this file.

Human says don't log, or they said they are done with this Skill / CentricMem → skip, do not accumulate.

MCP missing, librarian down, or token failed: do not drop the record. Compose the sweep this turn (named shelf id, note / decision / done bodies, this chat's transcript path). Persist it in this agent’s own memory titled CentricMem deferred sweep (no secrets, no keys). Send the authenticate link once (step 0.2). Keep working. Do not bootstrap, curl, or CLI-write. Closing the tab still does not file — the hold is so a later connected chat can.

When cm_health succeeds this chat: file this chat's Non-Micro batch, then file every CentricMem deferred sweep still in this agent’s memory (cm_keep those transcript paths from disk, then cards with attach). Same named-shelf rules. Do not ingest every transcript in the folder — only this chat and held sweeps.

Named shelf. Ambient never reports unclassified. Pick:

1. cwd linked (sourceDir match), or corpus=<slug> → that shelf
2. work belongs in an existing named shelf on libraries= → that id
3. none fit → default key: cm_library {id} (slug like my-project) then file there. Extra key, or cm_library missing: send the authenticate link (step 0.2) so they enter the default key to mint a shelf (Connect does not create shelves); hold the sweep in agent memory. Writes without a named shelf are 400 LIBRARY_REQUIRED. Do not treat workspace.current or an unmatched guest cwd as the shelf.

Label. If that id’s display name is a leftover slug and you know the product/human name, cm_library {id, displayName} this turn (REFERENCE). Do not mint a second shelf. Writes still shelf=<id>.

Then, with shelf= / library= that id:

1. Transcript → R2. If this runtime wrote a local transcript for this chat, Shell-read it; never paste it. cm_keep with that filename + file bytes (never path=). Leave the local file in place (do not delete chat state). Cursor: ~/.cursor/projects/<workspace>/agent-transcripts/<uuid>/<uuid>.jsonl. Claude Code, Codex, Kiro, Kilo, Copilot, and the rest: only if that agent actually wrote a file. No file → skip transcript keep; still file note / decision / done.
2. Then cm_note cm_log_decision cm_done with attach = that keep pointer. One sweep, one batch. If they keep talking, another sweep is OK for new facts — do not re-file the same decision.

Leftover unclassified on an old hub: dest must exist (cm_library if needed). cm_copy {from:unclassified,to:<named>} (identical skip; collisions imported/kept/from-unclassified/). Then cm_delete {id:unclassified}. Never copy to Inbox. Never download originals to this computer. Extra keys may copy if both grants; only default/login may delete a leftover shelf. Card delete stays login-only.

Organize leftover named shelves the same way: cm_copy {from,to} then cm_delete {id} — this deletes the leftover shelf (no restore warehouse). To curate a subset of cards onto another named shelf, cm_search / cm_show then cm_move {from,to,files}. Source cards are removed. Decision numbers stay if free on dest, otherwise a new seq. Attachments re-attach on dest. Default key or owner login. Extra keys cannot. Card delete stays login-only. Never download originals here.