ltm

# LTM — Memory Contract 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. ## Tools 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 | | `forget` | Remove a wrong or outdated memory.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_learn` belong to the OpenCode and Pi adapters — not the Claude server. On Claude, use the short names above. ## Ritual 1. **Recall before** — `recall` with natural language (FTS5 + semantic fallback), not bare keywords. Project restore → `context`. 2. **Trace** — for decisions with history, `graph` from the top recall hits. 3. **Work** — grounded in what recall returned. 4. **Learn after** — `learn` durable insights; `relate` to connect them. ## Categories `learn` takes one: **preference** (conventions/style) · **architecture** (design decisions) · **gotcha** (pitfalls) · **pattern** (reusable solutions) · **workflow** (process) · **constraint** (must-follow rules). ## Phase Map | 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`. ## Examples - Before coding → `recall "how we handle X"`, then `graph` the top hits. - Hit a non-obvious gotcha → `learn` it (category `gotcha`, importance 4); switching project → `context <project>`.