memory-management
How to extract durable memories from sources and write them as reviewable candidate memory files with provenance, confidence, and sensitivity.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
How to extract durable memories from sources and write them as reviewable candidate memory files with provenance, confidence, and sensitivity.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Required procedure for answering questions about the client - grounding, citation format, conflict handling, and how to say "not in the workspace".
How to record completed work in logs/agent-log.md so humans can inspect what the agent read, wrote, and concluded.
How to create deliverable artifacts (meeting prep briefs, draft emails, analyses) in artifacts/ with required structure and source manifests.
How to orient inside a client workspace - the directory layout, the manifest, provenance records, and which sources outrank which.
| name | memory-management |
| description | How to extract durable memories from sources and write them as reviewable candidate memory files with provenance, confidence, and sensitivity. |
Memories are stable facts, preferences, goals, relationships, and unresolved issues worth carrying across sessions. Agents NEVER write durable truth directly — they propose candidates that humans review.
raw observation -> memory/candidates/ (you write here) -> human review -> memory/durable/
You may only write into memory/candidates/. Never write to or edit memory/durable/.
What does NOT qualify: transient logistics ("meeting moved to Tuesday"), speculation, anything already captured in memory/durable/ (check first to avoid duplicates).
One memory per file. Path: memory/candidates/<YYYY-MM-DD>-<slug>.md (date = today). Required format:
---
id: <slug>
created: <ISO timestamp>
status: candidate
confidence: high | medium | low
sensitivity: public | internal | confidential | restricted
sources:
- <relative-path>:<line-or-range>
expires: <ISO date or "none">
supersedes: <path of older memory it replaces, or "none">
---
# <One-line memory statement>
<2-4 sentences of context: what was observed, when, and why it is durable.>
## Why this confidence level
<1-2 sentences justifying the confidence rating.>
sources must point at primary sources (transcripts, emails, client JSON) with line numbers.confidence: high requires an explicit, unambiguous statement in a primary source.confidence: low for inferences; state the inference chain.supersedes to that file and explain the conflict in the body — do not edit the durable memory.