Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use paxm as an active agent memory layer. Trigger when the user asks to recall prior context, search or inspect memory, remember working state or a durable fact, debug paxm history/metrics, or when a task would benefit from active memory recall before answering, especially repo, project, preference, architecture, or previous-decision questions.
.claude/skills/pax-beehive-paxm/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-08 | ✓→✓ | ＝ Same ✓ | -4% | 0% |
Use paxm as a supporting memory layer, not as a replacement for current source verification. Keep setup, credentials, provider selection, and passive hook policy under the user's control.
Check the local installation only when the task needs memory:
bashpaxm version paxm config doctor
If the binary is missing, ask the user before running the bundled installer. If the config is missing or invalid, ask whether to start the interactive setup flow. Do not edit paxm YAML by hand.
Use a short, concrete query and a small result limit:
bashpaxm recall --query "repo decision about hook ownership" --limit 3 --json
For a custom config, pass --config before the subcommand. Read scores and provider metadata, then verify facts that may have changed in the current repo. If a result exposes a precise lead, do at most two or three focused follow-up queries rather than running an open-ended search loop.
When the host exposes paxm MCP tools, prefer paxm_recall, paxm_remember, paxm_history, and paxm_config_doctor over shelling out.
Use STM for task-local working state:
bashpaxm remember --profile stm --text "Working note: ..."
Use LTM only for durable decisions, preferences, recurring fixes, or stable project conventions:
bashpaxm remember --profile ltm --text "Decision: ..."
Never store secrets, API keys, access tokens, raw private logs, or large pasted transcripts. Ask before storing sensitive personal or business information.
Passive Codex hooks are installed by this plugin only after the user completes the plugin-aware setup flow. Do not run a second paxm setup without the --integration codex-plugin mode, because that would change hook ownership.
When debugging, use bounded local history and logs:
bashpaxm history --days 7 --json paxm logs --tail 50
Treat memory as supporting evidence. Current repository files, current tool output, and explicit user instructions take precedence over recalled content.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,668 | 11,914 | +225% | 1 | 1 | 0% | 458 | 1,374 | +200% | 0 | 0 | — |
case-02 | fail→fail | 2,843 | 7,238 | +155% | 1 | 1 | 0% | 206 | 947 | +360% | 0 | 0 | — |
case-03 | fail→fail | 7,378 | 5,971 | -19% | 1 | 1 | 0% | 1,172 | 830 | -29% | 0 | 0 | — |
case-04 | fail→fail | 7,283 | 3,551 | -51% | 1 | 1 | 0% | 1,228 | 1,067 | -13% | 0 | 0 | — |
case-05 | fail→pass | 6,659 | 2,810 | -58% | 1 | 1 | 0% | 1,113 | 955 | -14% | 0 | 0 | — |
case-06 | fail→pass | 5,857 | 2,836 | -52% | 1 | 1 | 0% | 868 | 880 | +1% | 0 | 0 | — |
case-07 | fail→fail | 2,982 | 6,878 | +131% | 1 | 1 | 0% | 367 | 959 | +161% | 0 | 0 | — |
case-08 | pass→pass | 7,761 | 5,739 | -26% | 1 | 1 | 0% | 1,233 | 1,178 | -4% | 0 | 0 | — |
case-09 | pass→pass | 4,127 | 2,129 | -48% | 1 | 1 | 0% | 761 | 892 | +17% | 0 | 0 | — |
case-10 | pass→pass | 8,161 | 1,783 | -78% | 1 | 1 | 0% | 1,178 | 716 | -39% | 0 | 0 | — |
case-11 | pass→pass | 5,447 | 2,856 | -48% | 1 | 1 | 0% | 914 | 864 | -5% | 0 | 0 | — |
case-12 | fail→pass | 8,681 | 1,964 | -77% | 1 | 1 | 0% | 1,273 | 820 | -36% | 0 | 0 | — |
case-13 | pass→pass | 10,028 | 1,686 | -83% | 1 | 1 | 0% | 1,549 | 686 | -56% | 0 | 0 | — |
case-14 | pass→pass | 3,794 | 2,061 | -46% | 1 | 1 | 0% | 576 | 790 | +37% | 0 | 0 | — |
case-15 | fail→pass | 8,130 | 1,448 | -82% | 1 | 1 | 0% | 1,241 | 647 | -48% | 0 | 0 | — |
case-16 | pass→pass | 5,449 | 1,729 | -68% | 1 | 1 | 0% | 935 | 678 | -27% | 0 | 0 | — |
case-17 | fail→fail | 5,513 | 8,410 | +53% | 1 | 1 | 0% | 564 | 739 | +31% | 0 | 0 | — |
case-18 | pass→pass | 13,008 | 2,693 | -79% | 1 | 1 | 0% | 1,970 | 848 | -57% | 0 | 0 | — |
case-19 | pass→pass | 5,579 | 3,905 | -30% | 1 | 1 | 0% | 732 | 1,014 | +39% | 0 | 0 | — |
case-20 | pass→pass | 6,842 | 5,957 | -13% | 1 | 1 | 0% | 1,171 | 1,388 | +19% | 0 | 0 | — |
case-21 | pass→pass | 5,729 | 7,360 | +28% | 1 | 1 | 0% | 1,006 | 1,169 | +16% | 0 | 0 | — |
case-22 | pass→pass | 6,043 | 4,128 | -32% | 1 | 1 | 0% | 1,038 | 1,202 | +16% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +18 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.