claude-mem-value-eval — Mission Control

← Back to Tasks

Monday 09:00 SAST. Evaluates whether claude-mem adds value over the rest of Luci's memory stack (vault.db + SecondBrain, file auto-memory, MC-DB) — so we can decide if running both is worth it.

Runs claude (Sonnet) to do the judgement, writes a verdict to /tmp/claude-mem-eval.txt, and Telegrams Elmar the result. notify_on: always so the verdict lands every week, not just on failure.

Context: claude-mem was installed by Lucienne and runs side-by-side with the auto-memory layer (see memory claude-mem-sanctioned). This task is the recurring check that the side-by-side arrangement still earns its keep. runtime_profile: claude_sonnet_medium matches the actual claude CLI call.

Headless-hang hardening (MC-3772): the claude -p call uses --strict-mcp-config (boots zero MCP servers — the eval needs only Bash/Read/Grep/Glob, and MCP server spawns are the slowest, most memory-hungry part of boot on this OOM-pressured box), --no-session-persistence, and < /dev/null stdin. It is wrapped in timeout -k 30 1500 so a stalled claude is killed below the 1800s task ceiling and notify.py still reports the outcome instead of the whole task being reaped silently.

Schedule Mon at 09:000 9 * * 1

EnabledYes

Run Asshell

Timeout1800s

RetryNo

Next Runin 2d (2026-06-15 09:00)

Task Definition

tasks/claude-mem-value-eval.md

Run Asshell

Commandset

Tool-lessNo

BackgroundNo

Runtime profile

Notify destination

cd ~/workspace && timeout -k 30 1500 claude -p 'Weekly claude-mem value evaluation. Decide whether claude-mem still earns its place in the Luci memory stack. Steps: (1) Inspect claude-mem stored data and usage — look in ~/.claude-mem/ and note how much was captured in the last week and whether it is being queried. (2) Compare against the other memory layers: vault.db plus SecondBrain (knowledge base), file auto-memory at ~/.claude/projects/-home-lucienne-workspace/memory/, and the MC database work-state. (3) Decide whether claude-mem surfaces anything the other layers do not, and whether it is worth running both. Output a short verdict on the first line — exactly one of KEEP BOTH, CLAUDE-MEM REDUNDANT, or NEEDS REVIEW — then 3 to 5 lines of reasoning. Be concise.' --settings ~/.claude/settings-worker.json --allowedTools Bash,Read,Grep,Glob --strict-mcp-config --no-session-persistence < /dev/null > /tmp/claude-mem-eval.txt 2>&1; python3 ~/workspace/scripts/notify.py "Weekly claude-mem value eval:" "$(cat /tmp/claude-mem-eval.txt)"

This task calls claude directly. If the scheduler profile is GLM, Kimi, or MiniMax, prefer ~/workspace/PKA/scripts/claude-provider-run.sh so backup providers run in API-key-only mode with explicit settings/MCP context.

---
id: claude-mem-value-eval
name: Weekly claude-mem value evaluation
schedule: "0 9 * * 1"
enabled: true
criticality: low
timeout: 1800
retry: false
run_as: shell
notify_on: always
runtime_profile: claude_sonnet_medium
tags: [memory, evaluation]
command: |
  cd ~/workspace && timeout -k 30 1500 claude -p 'Weekly claude-mem value evaluation. Decide whether claude-mem still earns its place in the Luci memory stack. Steps: (1) Inspect claude-mem stored data and usage — look in ~/.claude-mem/ and note how much was captured in the last week and whether it is being queried. (2) Compare against the other memory layers: vault.db plus SecondBrain (knowledge base), file auto-memory at ~/.claude/projects/-home-lucienne-workspace/memory/, and the MC database work-state. (3) Decide whether claude-mem surfaces anything the other layers do not, and whether it is worth running both. Output a short verdict on the first line — exactly one of KEEP BOTH, CLAUDE-MEM REDUNDANT, or NEEDS REVIEW — then 3 to 5 lines of reasoning. Be concise.' --settings ~/.claude/settings-worker.json --allowedTools Bash,Read,Grep,Glob --strict-mcp-config --no-session-persistence < /dev/null > /tmp/claude-mem-eval.txt 2>&1; python3 ~/workspace/scripts/notify.py "Weekly claude-mem value eval:" "$(cat /tmp/claude-mem-eval.txt)"
---

Monday 09:00 SAST. Evaluates whether claude-mem adds value over the rest of
Luci's memory stack (vault.db + SecondBrain, file auto-memory, MC-DB) — so we
can decide if running both is worth it.

Runs `claude` (Sonnet) to do the judgement, writes a verdict to
`/tmp/claude-mem-eval.txt`, and Telegrams Elmar the result. `notify_on: always`
so the verdict lands every week, not just on failure.

Context: claude-mem was installed by Lucienne and runs side-by-side with the
auto-memory layer (see memory `claude-mem-sanctioned`). This task is the
recurring check that the side-by-side arrangement still earns its keep.
`runtime_profile: claude_sonnet_medium` matches the actual `claude` CLI call.

Headless-hang hardening (MC-3772): the `claude -p` call uses `--strict-mcp-config`
(boots zero MCP servers — the eval needs only Bash/Read/Grep/Glob, and MCP server
spawns are the slowest, most memory-hungry part of boot on this OOM-pressured
box), `--no-session-persistence`, and `< /dev/null` stdin. It is wrapped in
`timeout -k 30 1500` so a stalled `claude` is killed below the 1800s task ceiling
and `notify.py` still reports the outcome instead of the whole task being reaped
silently.

Edit Schedule

Frequency

Result:

Runs (7d)

Completed

Failed

Timeout

100%

Success (7d)

2.3m

Avg Duration

memory evaluation

Run History (1)

Started	Finished	Status	Duration	Output	Log
5d ago	5d ago	completed	2.3m	Telegram send failed: HTTP Error 400: Bad Request Sent	View log

Comments (0)

No comments yet — add instructions for Luci

←