メニュー
The LTM memory-tool contract — names, ritual, categories, phase map. USE WHEN recalling, learning, relating, or restoring memory.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
USE WHEN separating trusted instructions from untrusted content or tool output.
USE WHEN hardening command execution, path handling, or installer boundaries.
USE WHEN handling credentials, env vars, logs, tests, or docs safely.
Reference for LTM memory commands, context hooks, and DB schema; use when the user mentions learning, recalling, forgetting, relating, or capturing memory, or when a session starts or the project changes.
Mines LTM memories from past git commits. Use when onboarding a repo into LTM, backfilling history after enabling gitLearn, or harvesting patterns after a sprint.
Reference for retrieved patterns and lessons from past sessions; use when prior fixes, gotchas, decisions, or earlier sessions may apply, or when asked what we learned before.
| name | Ltm |
| description | The LTM memory-tool contract — names, ritual, categories, phase map. USE WHEN recalling, learning, relating, or restoring memory. |
Single source of truth for talking to long-term memory. Load it whenever you touch memory so tool names and the recall-before / learn-after ritual stay consistent.
Plugin ltm, server key memory. In Claude Code each is called as mcp__plugin_openltm_memory__<name>.
| Tool | Call when |
|---|---|
recall | Before any non-trivial task — surface prior decisions, gotchas, patterns. |
learn | After a durable pattern, gotcha, or decision. |
context | Session start / project switch — restore goals, decisions, gotchas. |
context_items | List one type (goal, decision, progress, gotcha) for a project. |
graph | Trace decision chains — pass seed IDs from recall to see the "why". |
relate | Link two memories (supports, contradicts, refines, depends_on, related_to, supersedes). |
forget | Remove a wrong or outdated memory. |
admin_audit | Inspect a memory's provenance / audit trail. |
ltm_recall/ltm_learnbelong to the OpenCode and Pi adapters — not the Claude server. On Claude, use the short names above.
recall with natural language (FTS5 + semantic fallback), not bare keywords. Project restore → context.graph from the top recall hits.learn durable insights; relate to connect them.learn takes one: preference (conventions/style) · architecture (design decisions) · gotcha (pitfalls) · pattern (reusable solutions) · workflow (process) · constraint (must-follow rules).
| Phase | Before | During | After |
|---|---|---|---|
| Spec | recall · context | explore code | learn · relate |
| Plan | recall (main thread) | graph chains; optional reasoning API* | relate |
* Planner has no MCP tools — main thread recalls and injects a ### Pre-Plan Context block. Richer chains when the reasoning server runs: GET http://localhost:7331/api/reasoning/search?q=<topic>&depth=2.
recall "how we handle X", then graph the top hits.learn it (category gotcha, importance 4); switching project → context <project>.