Browse

Delete a memory entry by key. Use to remove outdated or incorrect facts from persistent memory.

Run a relational query over agent state. Relations: `memory` (persistent memory entries: key, value, namespace, type, references, createdAt, updatedAt), `tool_calls` (recent tool call history: name, args, result, timestamp), `tasks` (active goal's task list: id, desc, status, rounds, note). Query language is expr-lang: filter(relation, predicate) selects rows (e.g. filter(memory, .type == "error")), map(relation, expr) projects fields, join(left, right, leftField, rightField) equi-joins two relations into rows with left_/right_ prefixed fields, union(a, b) combines two relations with set semantics.

Save a fact to persistent memory. The memory is injected into every LLM call automatically. Use for project conventions, user preferences, key decisions, or any information that should persist across sessions. Keys should be short and descriptive.

Search persistent memory for entries matching a query. Returns up to 20 matching key-value pairs (each value capped), best match first: a key hit outranks a value-only hit, more query words outrank fewer, and a newer update breaks a tie. Use to recall previously saved facts, preferences, or decisions. There is no memory_list tool; the memory graph is already in the system prompt.

Extract text from an image via tesseract OCR. Runs entirely locally -- no API key required. Supports png, jpg/jpeg, tiff, bmp, gif. Requires: file (absolute path to the image). Optional: language (tesseract -l value, e.g. 'eng' or 'eng+fra'; default 'eng'). Does not support PDF -- use load_document with type pdf for text-layer PDFs.

Search the web using Perplexity AI. Provides cited, up-to-date answers from the internet. Requires PERPLEXITY_API_KEY. Input is a plain search query or question.

Read a file from the local filesystem. Returns the contents with a 1-based line number on every line (` 42\tcode$`) - use those numbers for edit_file insert/view. Tabs render as ^I, other control characters as ^X, and a $ marks the true end of each line, so indentation, trailing whitespace, and stray control characters are visible instead of hidden. IMPORTANT: ^I, ^X, and the trailing $ are a display convention this tool adds - the real file does NOT contain those literal characters. When quoting text back into edit_file's old_str, write the real tab/control character (or match a whole line via view_range/anchor instead), never the literal caret sequence - copying "^I" as text will not match a real tab byte. Plain text, source code, configuration files, and documentation are read directly; PDF, DOCX, EPUB, RTF, and ODT documents have their text extracted automatically. Use load_document instead for CSV/HTML structured parsing or RAG chunking. Pass match_id (from a search_local result) instead of file_path/offset to jump straight to that hit.