claude-cowork-agentops

A Claude Cowork automation kit that runs a daily housekeeping pass over claude-mem and distills it into conventions (CLAUDE.md), reusable logic (skills), and long-term memory (Obsidian) — with AgentOps (monitoring, alerting, evaluation, governance) built in.

claude-mem is a recency-window local memory: it carries recent context into the next session, and older knowledge gets pushed out and lost. This kit is the daily batch that rescues what should persist into the long-term layer before it ages out.

Set up entirely from Cowork chat: paste setup.md and Claude will ask you a few questions (output language, project folder, vault, which tasks) and configure everything for you.

The daily monitoring dashboard in Obsidian (Dataview + Charts): per-machine activity/token table and trend lines, generated automatically each night.

Memory hierarchy

Layer	Role	Location
CLAUDE.md	Conventions / structure (per project)	each project (proposed uncommitted in git)
skill	Reusable logic	project-specific → <project>/.claude/skills/ (auto, uncommitted); cross-project → ~/.claude/skills/ (proposal)
claude-mem	Short/mid-term memory (local, transient)	~/.claude-mem
Obsidian vault	Long-term memory (shared across devices)	local path (synced via Obsidian Sync, etc.)

Flow

Claude Code work -> claude-mem (local, short/mid-term)
        | daily batch (Cowork scheduled task / 00:00)
        |-> CLAUDE.md   : conventions/structure (proposed uncommitted in the most relevant dir)
        |-> skills      : project-specific -> <project>/.claude/skills (auto); cross-project -> ~/.claude (proposal)
        \-> Obsidian vault (long-term)
              |- reports/YYYY-MM/DD/<machine>.md      : that day's work log
              |- knowledge/<slug>.md                   : extracted reusable knowledge (atomic notes)
              |- monitoring/YYYY-MM/DD/<machine>.md    : claude-mem metrics
              \- adoption/YYYY-MM/DD/<machine>.md      : proposal acceptance rate

Components (scripts, stdlib-only Python 3)

Script	Role
memory_digest.py	Extract incremental memory from claude-mem.db (state stored locally under ~/.claude-mem)
monitoring_digest.py	Daily activity / tokens / memory growth / health note (TZ-correct, real-timestamp log aggregation)
adoption_eval.py	Acceptance rate of CLAUDE.md proposals (committed vs pending) from git
hotspot.py	Files with repeated bugfixes over N days = tech-debt hotspots
knowledge_audit.py	Surface duplicate / stale knowledge notes (helper for monthly consolidation)
redact.py	Mask API keys / tokens / private keys before writing to the vault
health_check.py	Inspect claude-mem worker / errors / WAL and judge OK/WARN/ALERT
backfill.py	(optional, one-time) Backfill monitoring/adoption for past dates if you already had claude-mem — pick the period

All scripts read the claude-mem DB from a copy (never touch the live WAL). Machine name and timezone come from ~/.claude-mem/machine.json (local, not synced). Script output is English; the LLM-authored reports/knowledge use your configured language.

Scheduled tasks (Cowork)

taskId	cron	Role
claude-mem-housekeeping	0 0 * * *	Daily housekeeping (previous day). Apply CLAUDE.md + write reports/knowledge/monitoring/adoption
claude-mem-healthcheck	0 */2 * * *	Health check every 2h. Alert to Slack + Obsidian on WARN/ALERT only
claude-mem-hotspot	0 9 * * 1	Weekly tech-debt hotspot detection
claude-mem-knowledge-consolidate	0 9 1 * *	Monthly knowledge merge/update/prune (run on ONE machine only)
claude-cowork-agentops-update	0 23 * * *	(optional) Check-only update detector. Notifies Slack when the repo is behind; never pulls

Task prompts to register are in scheduled-tasks.md.

Keeping the batch up to date

The scheduled task above only checks for updates — it runs scripts/check_update.sh, which compares the remote branch tip against the local HEAD with git ls-remote. That is a read-only network query: it never writes to .git, so it cannot create the stale lock files that make VM-side git pull fragile. When an update is available it notifies Slack; you then pull on the host (git pull --ff-only) and re-register any scheduled tasks if scheduled-tasks.md changed.

AgentOps pillars

• Observability — daily monitoring metrics + Dataview/Charts dashboards.
• Reliability (active alerting) — health check detects outages (worker down, disk I/O, WAL anomalies) and notifies Slack.
• Feedback (evaluation) — adoption rate quantifies whether proposals are useful (provenance marker isolates pipeline-generated CLAUDE.md).
• Governance — redact keeps secrets out of the cloud-synced vault.

Requirements

• claude-mem (~/.claude-mem/claude-mem.db)
• Claude Cowork (scheduled tasks + folder access)
• Obsidian + plugins Dataview (enable JavaScript Queries) and Obsidian Charts
• Python 3 (standard library only)

Quick start

Paste setup.md into Cowork chat and follow the guided setup, or do it manually per INSTALL.md. In short:

1. git clone this repo directly inside your projects root (the dev folder you connect to Cowork).
2. Create ~/.claude-mem/machine.json: {"machine":"<name>","utc_offset_hours":<offset>}.
3. Connect ~/.claude-mem, your projects root, and a local Obsidian vault in Cowork.
4. Register the scheduled tasks from scheduled-tasks.md (choose your report/knowledge language).
5. For Slack alerts: create a channel, put its channel_id in the health-check prompt, Run now once to approve.

Excluding directories from writes (optional)

To stop housekeeping from touching CLAUDE.md / SKILL.md in certain repos, create ~/.claude-mem/housekeeping-config.json:

{
  "exclude_write_paths": [
    "org/some-repo",
    "/absolute/path/to/another"
  ]
}

Anything under a listed directory is never created or updated — no CLAUDE.md, no project skills. Those projects are still read, analyzed, written up in the daily report, and mined for knowledge/ notes; only the writes are withheld, and each withheld write is logged under "Skipped by exclusion config" in the report. Matching is path-boundary prefix; entries may be relative to the projects root or host-absolute. No file = no exclusions.

Backfill (optional — existing claude-mem users)

Already had claude-mem before installing this? You can backfill past dates. The range can be large, so you pick the period.

python3 backfill.py --list-range                          # see how far back memory goes
python3 backfill.py --vault "<vault>" --days 14            # monitoring + adoption for last 14 days
python3 backfill.py --vault "<vault>" --from 2026-05-01 --to 2026-05-31

monitoring/ and adoption/ are deterministic (instant). reports/ and knowledge/ are LLM-generated, so backfill them via the claude-mem-backfill task in scheduled-tasks.md — keep that window small. Backfill never commits or moves the daily incremental checkpoint.

Design notes

• Keep the vault on a local path (cloud streaming paths are not mounted in Cowork task VMs). Sync across devices with Obsidian Sync or git.
• reports/monitoring/adoption are namespaced by <machine>.md, so multiple machines can write to the same vault without conflict. knowledge is shared — run monthly consolidation on one machine only.
• CLAUDE.md is proposed uncommitted. Review with git diff; commit = accept, checkout = discard.
• CLAUDE.md placement is hierarchical: the root CLAUDE.md stays high-level (overview, conventions, project settings, environment, architecture); module-specific detail / file-level specifics / localized gotchas go into the CLAUDE.md of the relevant subdirectory (package/module boundary or where changes concentrate).
• Skills are placed by scope: a skill specific to one project is written to that project's <project>/.claude/skills/ (auto, uncommitted — review like CLAUDE.md); a cross-project skill targets ~/.claude/skills/.
• ~/.claude/ (global CLAUDE.md / cross-project skills) is not writable from Cowork, so those are proposals only (apply manually).

License

MIT — see LICENSE.