context-loader

Files (1)

• SKILL.md

SKILL.md (first lines)

# Context loader

The session-start hook already loads memory once at the top. This is the manual version for mid-session, when the topic shifts to something the startup recall didn't cover.

When the topic maps to a repo or project, first call `get_memories` with `facets: true` — it returns the store's real vocabulary (which `project` tags, `type` values and their counts exist) instead of items. Use a discovered `project` value to scope the searches via `filters: {"project": "<value>"}`; guessing tag spellings loses to looking them up. If the call errors (server without the facets patch), skip this step silently.

Then take the current topic — from the user's phrasing or the files in play — and run two or three `search_memories` calls from different angles rather than one broad query: the topic itself, the topic plus "decision", the topic plus "gotcha"/"doesn't work". Use `rerank=true` and `exclude_sources: ["session", "archived"]` so raw capture chunks don't crowd out facts.

Merge by id, keep the strongest handful, and surface them compactly:

```
Loaded <n> memories on "<topic>":
- [decision] ... [mem0:id]
- [anti_pattern] ... [mem0:id]
```

If nothing relevant comes back, stay silent — don't announce an empty load.

From pjope/mem0-selfhosted-hooks, MIT.

More from pjope/mem0-selfhosted-hooks