Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/acn-ericlaw/agent-memory/refresh-metadatanpx skills add acn-ericlaw/agent-memory --skill refresh-metadatagit clone --depth 1 https://github.com/acn-ericlaw/agent-memoryWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00075 | $0.00922 |
| Opus 5 | $0.00037 | $0.00461 |
| Sonnet 5 | $0.00015 | $0.00184 |
| Haiku 4.5 | $0.00007 | $0.00092 |
Grade A, and why
refresh-metadata scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 57 lines — stays where its author put it; the contents beside it link to each section on GitHub.
⚠️ Tool-managed skill provided by agent-memory (
provenance: agent-memory-builtin). Don't edit it in place — fork under a new name, or upstream a fix to the agent-memory project (seeSKILLS.md).
This skill performs REVIEW.md steps 2–3 (apply events + re-tier) as a runnable script. For every
fact in continuity.md it recomputes last_used, uses, and tier from the ## Memory References across
memory/sessions/, and writes the footers back. This is the "full rebuild" path REVIEW.md already
calls "deterministic and reproducible by any agent" — pure arithmetic, no judgment — so it's safe to
mechanize. Agents routinely skip this pass (they archive faded facts but don't re-tier the ones that
stay), leaving stale footers; a cross-vendor review proved it. This closes that gap.
What it does NOT do
- It does not decide what to archive — that's
archive-factplus the agent's judgment (never-pick-a-winner). It clamps tier atarchive-candidate; a fact still in continuity is never markedarchived(that tier means moved). - It does not touch
coreorsupersededfacts (human-set / terminal), or facts with no reference in any session log (legacy — can't recompute, so preserved), and it never adds missing fields. - For a pinned
- [ ]open thread it refreshes the factual fields (uses/last_used) but leaves the tier label as-is (v4.26.1) — pinned-ness protects an open thread, not its tier, so the tool doesn't opine on it (aworking-tagged open thread is fine).
So the division of labor across the three memory tools mirrors the meaning/mechanics split:
memory-lint verifies (read-only) · refresh-metadata re-tiers (arithmetic) · archive-fact
moves (the decided archival). Only which facts to archive needs the agent.
How to run
From the repo root, whichever runtime the machine has (output is byte-identical):
python3 agent-skills/refresh-metadata/scripts/refresh-metadata.py [--dry-run]
# or
node agent-skills/refresh-metadata/scripts/refresh-metadata.mjs [--dry-run]
What ships with it
4 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 57 lines · 75 tokens per session scan A f42a07e7bbda
refresh-metadata is a skill published in the GitHub repository acn-ericlaw/agent-memory (5 stars, last pushed 8d ago), licensed Apache-2.0. It adds 75 tokens to every session and 922 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
reelier-slice
Use when implementing any change to the Reelier authority ledger or Path C compiled-authority code - encodes the evidence-led slice loop, the baseline gate, the RED/GREEN review protocol, and the failure modes that have actually cost slices here.
reelier-replay
Freeze a repeatable, tool-call-driven task into a replayable Reelier skill, and replay an existing skill instead of redoing deterministic work. Use after finishing a task that was mostly API/MCP tool calls (data pulls, report generation, deploy checks, CRUD sequences) — never for arbitrary coding/file-edit sessions…
reelier-write-safety
Bound and record an agent's writes before granting them. Use when the user is adding or configuring an MCP server, planning to run an agent unattended or on a schedule or in CI, or deciding whether to give an agent write access to a real system. Covers the recorder, the policy seatbelt, approvals, and what a receipt…
reelier
Replay recurring deterministic jobs at 0 tokens instead of re-reasoning them every run. For cron/heartbeat tool-call workflows (status checks, data pulls, report generation, CRUD sequences) — record the job once with reelier, then run "npx -y reelier run " on every heartbeat and gate drift with "reelier diff". Never…
coding
编写并运行 Python 代码，验证脚本逻辑和输出。.
kayba-stage-6-hitl
Human-In-The-Loop gate that presents the action plan with full context, collects an informed approval/modification/rejection decision, and records the outcome. Trigger when the user says "run stage 6", "HITL review", "approve action plan", or when invoked by the kayba-pipeline orchestrator. Requires eval/actionplan.md…