Tools Reference

Tools Reference

🇬🇧 English

The daemon exposes these tools over MCP (to AI clients) and REST (POST /api/v1/<tool>). All are defined in tool-handlers.ts (MEMORY_TOOL_DEFS).

recall

Search the vault. Step 1 of a two-step flow — returns lean candidates, no bodies. Call load_memory only for the hits you actually need.

Param	Type	Default	Notes
query	string	—	Natural-language query or a description of what you're about to do. Required.
k	number	5	Max results (1–20).
scope	string	—	Exact-match filter, e.g. carnexus, user-preference.
type	string	—	Exact-match filter, e.g. lesson, decision, project-fact.
verbosity	"lean" | "full"	"lean"	full adds matched_terms, mode, hop, topic_path, the stages block, and the untruncated summary.
min_score	number	30	Drop hits below this score (the noise floor). Raise for high-confidence only.

Lean hit: { id, title, type, scope, summary, score } (summary ~160 chars). Full hit additionally: { matched_terms, mode, hop, topic_path } + full summary.

load_memory

Load one memory's content by id. Step 2 of the recall flow.

Param	Type	Default	Notes
id	string	—	The memory slug (no .md). Required.
verbosity	"lean" | "full"	"lean"	lean = essential frontmatter + body without the auto-related block. full = complete frontmatter (related_via cosines, source, confidence, …) + raw body.

Lean frontmatter keys: id, title, type, scope, summary, topic_path, tags, recall_when, related, created, updated.

save_memory

Persist a memory as a Markdown file with YAML frontmatter. Save autonomously on a memory-worthy moment (correction, durable rule, finalized decision); don't wait to be asked.

Key fields: title, type (lesson | preference | user-preference | project-fact | decision | workflow | meta-working | reference), summary (≤400 chars), body, topic_path, tags, scope, and recall_when — 2–4 concrete trigger phrases, the highest-weighted search field. overwrite: true updates an existing id instead of erroring.

find_document / read_document

Search and read documents in the vault (PDFs, scans with OCR), separate from memories.

---

🇩🇪 Deutsch

Der Daemon stellt diese Tools über MCP (für AI-Clients) und REST (POST /api/v1/<tool>) bereit. Alle sind in tool-handlers.ts (MEMORY_TOOL_DEFS) definiert.

recall

Vault durchsuchen. Schritt 1 eines Zwei-Schritt-Flows — liefert lean Kandidaten, keine Bodies. load_memory nur für die Hits aufrufen, die du wirklich brauchst.

Param	Typ	Default	Hinweise
query	string	—	Natural-Language-Query oder Beschreibung dessen, was du gleich tust. Pflicht.
k	number	5	Max. Ergebnisse (1–20).
scope	string	—	Exact-Match-Filter, z.B. carnexus, user-preference.
type	string	—	Exact-Match-Filter, z.B. lesson, decision, project-fact.
verbosity	"lean" | "full"	"lean"	full ergänzt matched_terms, mode, hop, topic_path, den stages-Block und die ungekürzte summary.
min_score	number	30	Hits unter diesem Score droppen (Noise-Floor). Höher = nur High-Confidence.

Lean-Hit: { id, title, type, scope, summary, score } (summary ~160 Zeichen). Full-Hit zusätzlich: { matched_terms, mode, hop, topic_path } + volle summary.

load_memory

Inhalt einer Memory per id laden. Schritt 2 des Recall-Flows.

Param	Typ	Default	Hinweise
id	string	—	Der Memory-Slug (ohne .md). Pflicht.
verbosity	"lean" | "full"	"lean"	lean = essenzielle Frontmatter + body ohne Auto-Related-Block. full = komplette Frontmatter (related_via-Cosines, source, confidence, …) + roher body.

Lean-Frontmatter-Keys: id, title, type, scope, summary, topic_path, tags, recall_when, related, created, updated.

save_memory

Eine Memory als Markdown-File mit YAML-Frontmatter persistieren. Autonom speichern bei einem memory-würdigen Moment (Korrektur, dauerhafte Regel, finale Entscheidung); nicht auf Aufforderung warten.

Kernfelder: title, type (lesson | preference | user-preference | project-fact | decision | workflow | meta-working | reference), summary (≤400 Zeichen), body, topic_path, tags, scope, und recall_when — 2–4 konkrete Trigger-Phrasen, das höchstgewichtete Suchfeld. overwrite: true aktualisiert eine bestehende id statt zu erroren.

find_document / read_document

Dokumente im Vault durchsuchen und lesen (PDFs, Scans mit OCR), getrennt von Memories.