setup

Basic Memory setup

Run a short, adaptive interview (~2-3 minutes) and then write the configuration. Be conversational and skip questions whose answer is already obvious from context (e.g. if list_memory_projects shows a single local project and no cloud workspaces, don't ask about cloud/teams — just confirm). Suggest a sensible default for every question so the user can accept with one word. Don't do any writes until the interview is done and you've confirmed the plan.

Prerequisite check

First confirm the Basic Memory MCP server is connected — call list_memory_projects. If that tool isn't available or errors, Basic Memory isn't wired into Claude Code yet. Stop and walk the user through it first (everything below depends on it):

1. Install it: uv tool install basic-memory (or pip install basic-memory), version >= 0.19.0.
2. Connect it: claude mcp add basic-memory -- uvx basic-memory mcp, then restart the session so the MCP server loads.

Re-check list_memory_projects before continuing — don't start the interview until it succeeds.

Interview

Ask only what you can't infer. Cover:

1. Focus / how you'll use it. "What will this project mostly be — code/dev, research, writing, knowledge capture, planning, or a mix?" This answer is load-bearing, not small talk: it drives the folder structure you suggest (step 4) and is stored so the SessionStart brief can surface it, keeping capture matched to the use-case. Don't let it evaporate — if you infer it from context instead of asking, still say the use-case you assumed and the structure it implies, and let the user correct it in one word.

2. Project mapping. "Do you already have a Basic Memory project for this, or should I create one?"

   • Existing → show list_memory_projects() and let them pick. That name becomes primaryProject.
   • New → propose a name (default: this repo's directory name) and create it with create_memory_project.
     • Local project (default): path defaults to ~/basic-memory/<name>/; any connected Basic Memory server can create it.
     • Cloud project (the user wants capture in a cloud workspace): pass the workspace selector (a slug from list_workspaces) and a cloud-style path like /<name>, and create it with a cloud-connected MCP server. A purely local server (uvx basic-memory mcp) treats the path as a local directory and fails to create it (e.g. read-only /). When both a local and a cloud server are connected, route creation and the schema seeding through the cloud one, and pin primaryProject to the new project's external_id UUID (collision-proof across workspaces).

3. Cloud / teams (skip if there are no extra workspaces). Run list_workspaces. If the user belongs to more than one workspace, they likely have a team workspace alongside their personal/default one. Use list_memory_projects to see the projects in each (note: project names collide across workspaces, so always use the workspace-qualified name, e.g. my-team-2/notes, or the external_id UUID — never a bare name).

   • Read from the team (recommended): ask which team projects to pull into the session brief for recall. Store their qualified names in secondaryProjects. These are read-only — recall reads across them; nothing is written to them. Cap: the SessionStart brief reads only the first 6 shared projects per session (a latency/output bound), in list order. If the user wants more than six, order the most relevant first and tell them the rest are configured but not read each session.
   • Share target (optional): if the user wants a place to publish notes to the team via /basic-memory:share, add it to teamProjects as "<qualified-name>": { "promoteFolder": "shared" }. Sharing is always a manual gesture — auto-capture never writes to a team project.

   Keep primaryProject a project the user owns for their own capture; team projects are for reading and deliberate sharing only.

4. Placement — learn or suggest (depends on the project's state). The goal is a short placementConventions string (3-6 lines) telling you where new notes go. How you get it depends on whether the project already has notes:

   • Existing project with notes → learn. Inspect it: list_directory for the folder layout, sample a few notes per folder (and, where a folder holds recurring typed notes, you may run schema_infer to see their shape). Summarize the real conventions — folder-by-topic layout, naming style, the observation categories they favor. Infer from their actual notes; don't impose.
   • New or empty project → suggest (there's nothing to learn yet). Propose a light structure that fits the focus from step 1 — 3-5 optional top-level folders, no deep taxonomy — and be explicit that it's a starting point, not a scaffold: notes work fine without it and structure can stay emergent. Don't create empty folders; folders appear as notes land in them. Let the user edit or decline in one word. Either way, keep it short and store the result as placementConventions. The SessionStart brief surfaces it (alongside captureFolder), so this is what makes your captures land where the user expects — without it, placement is guesswork.

5. Schemas. "I'll add schemas for session checkpoints, decisions, and tasks so I can find them precisely later — okay?" (See "Seed the schemas" below.)

6. How active should I be? (output style) "Want me to proactively capture — search the graph before recalling, write material decisions as typed notes, and cite permalinks? Or keep it quiet (just the session brief, the PreCompact checkpoint, and /basic-memory:remember on demand)?" Enabling it sets outputStyle: "basic-memory". Default to enabled; leave it off for a recall-only, low-noise setup. (This is the single knob for how proactive the assistant is — the hooks always run regardless.)

7. Shared skills (optional, default yes). "Want the full Basic Memory toolkit — the shared memory-* skills (memory-notes, memory-tasks, memory-research, memory-schema, memory-defrag, …)? I can install them alongside this plugin." These are the canonical, framework-agnostic skills (the same set OpenClaw bundles). This plugin ships only the Claude-Code-specific glue and pulls the shared set on demand — it doesn't vendor its own copies. (See "Install the shared skills" below.)

Apply (after confirming the plan)

1. Seed the schemas

The plugin ships seed schemas at <plugin>/schemas/ — that's two directories up from this skill's directory, then schemas/ (this skill is at <plugin>/skills/setup/). Read session.md, decision.md, and task.md there.

For each one:

• Check whether the chosen project already has a schema for that type (search_notes with metadata_filters={"type": "schema"}, or try read_note("schemas/<name>")). If it exists, skip it — never overwrite a schema the user may have customized.
• Otherwise write it with write_note, routed to primaryProject (pass it as project, or as project_id if it's an external_id UUID):
  • directory="schemas", note_type="schema", title = the schema's title (Session / Decision / Task).
  • content = the markdown body only — everything after the --- frontmatter block (the # Session heading and the prose).
  • metadata = the schema's structured frontmatter as a nested dict: entity, version, the full schema map, and settings (keep its nested frontmatter, and pass enum values as JSON arrays, e.g. ["open","resumed","closed"]).
  • Do not put the schema's --- frontmatter inside content. On the cloud write path that nested YAML is silently coerced to the string '[object Object]' (basic-memory-cloud#1000), corrupting schema/settings. The metadata param round-trips correctly on both local and cloud. After seeding, verify one note with read_note(..., output_format="json", include_frontmatter=true) — schema/settings must come back as nested objects, not strings.

2. Install the shared skills (if the user opted in)

First, guard against clobbering a source checkout. If ./skills already exists, is tracked in git, and holds memory-* directories, you're inside the skills' own source repo (e.g. basic-memory itself) — the install would overwrite the working copy with published versions. In that case skip the install and tell the user the skills are already present as source; don't run the command. Quick check:

git ls-files skills/ | grep -q memory- && echo "source repo - skip install"

Otherwise, run from the project root:

npx skills add basicmachines-co/basic-memory --path skills

This installs the canonical memory-* skills into the user's skills directory — the single source of truth, shared with OpenClaw. The plugin does not vendor copies; it relies on this shared set. If npx / the skills CLI isn't available, point the user at the manual install in the top-level skills/README.md.

3. Write settings

Build the basicMemory block from the interview:

{
  "basicMemory": {
    "primaryProject": "<chosen>",
    "secondaryProjects": [],
    "captureFolder": "sessions",
    "rememberFolder": "bm-remember",
    "recallTimeframe": "3d",
    "preCompactCapture": "extractive",
    "placementConventions": "<learned or suggested summary, or null>",
    "teamProjects": {}
  },
  "outputStyle": "basic-memory"
}

Only include outputStyle if the user opted in. Ask whether this is a team default (write/merge into .claude/settings.json, suggest committing it) or personal (.claude/settings.local.json). Merge into any existing file — read it, add/replace only the keys above, preserve everything else. Use compact, valid JSON.

Writing the basicMemory block is also what stops the SessionStart hook's first-run nudge — the config's presence is the signal that setup has run.

4. Smoke-test the wiring

Before you close, prove recall actually resolves — this catches a misnamed project, missing cloud credentials, or a ref that doesn't route, while the user is still here to fix it. Run the same structured query the SessionStart hook runs, via the CLI it uses (basic-memory / bm / uvx basic-memory):

• Primary: … tool search-notes --type schema --page-size 5 against primaryProject — use --project-id <uuid> for a UUID, --project <ref> otherwise. It should return the three schemas you just seeded.
• One shared project (only if secondaryProjects is non-empty): a --type decision --status open query against the first ref. It just needs to return cleanly — 0 results is fine; an error means the ref doesn't route.

If a query errors, or the primary returns nothing, surface it and fix the project ref before closing — don't let the next session's brief come up empty.

Close

Confirm what you did in a few lines: the project mapping, which schemas were seeded vs. already present, whether placement was learned or suggested, the smoke-test result, and whether the output style is on.

Then handle activation based on the output style:

• Output style enabled → it's fixed at session start, so the full capture reflexes only take effect next session. Prompt the user to restart the session (start a new Claude Code session) to activate them. Be precise so it doesn't read as "nothing works yet": recall is already live this session (the SessionStart hook's prompt ran), and the PreCompact checkpoint works now too — only the proactive-capture reflexes wait for the restart.
• Output style off → no restart needed; the hooks already run.

End with: "Done — I'll use this from the next message. Run /basic-memory:status anytime to see what I'm tracking."