Store a new memory.

Return (creating if absent) the per-chat mutex, updating last-access time. Sweeps expired entries before inserting to keep the map bounded.

Execute a tool call from the LLM.

ASCII case-insensitive substring check for pre-lowercase trigger gates.

Convenience constructors keep the table readable.

Build a `Send` invocation handle for this skill. Synchronous and cheap: callers can hold the skill-loader lock across this, then drop it before `await

Append a message to a conversation.

Spawn `tegrastats` as a child process, parse each line, and broadcast the latest snapshot via a watch channel. Returns a receiver that always holds t

Set the memory store for memory tools (recall, forget, store).

Append a message, logging SQLite/IO failures instead of silently dropping them.

Get the raw JSON value for an argument.

Extract facts from a user utterance using pattern matching (Tier 1). Returns zero or more facts with categories: - `identity`: name, age, occupation,

Search memories with temporal decay applied. Returns results ranked by: BM25 relevance * temporal_decay_multiplier. Evergreen memories are exempt fro

Run a closure against the shared memory store on the blocking thread pool. genie-core's HTTP server and Telegram adapter share one OS thread (`tokio:

Decide whether a memory is safe to use in the current response context.

ASCII case-insensitive substring check for pre-lowercase trigger gates.

Create a new conversation. Returns the conversation ID.

Stable lowercase representation for logs/audit trails — independent of the derived `Debug` repr, which is free to change with variant names.

Voice output formatter. LLMs output markdown, bullet points, long paragraphs, and special characters that sound terrible when spoken by TTS. This mod

Mark a memory as promoted (moved to permanent storage). Convenience wrapper around [`mark_promoted_many`](Self::mark_promoted_many) for the single-id

Sanitize LLM output — remove any leaked secrets before showing to user. Scans for patterns that look like API keys, tokens, or credentials in the LLM

Return a freshly-created unique parent dir and a `memory.db` path inside it. `Memory::open` derives `canonical_dir = path.parent().join("memory")` and

All available tool definitions (for the LLM system prompt).

True if no trusted keys are loaded — every verification fails closed.

Build the system prompt with tools and memory context.

Build the memory section to append to the system prompt for a given query. Strategy: 1. Always include identity memories 2. Search for query-relevant

Load a single skill from a `.so` file.

Get recent N messages (for context window).

Read and parse one HTTP request from `reader`, enforcing every bound in `limits`. The entire read runs under a single deadline, so a stalled peer cann

Extract facts and store them, with deduplication. Returns the number of new memories stored.

Infer V1 policy metadata from the memory kind and content. This is used both when new memories are stored and when older databases are backfilled int

Force JSON output (any valid JSON).

Acquire the chat turn lock, bounded by `busy_wait`. Returns `None` if a turn is already in flight and does not release within the budget; the caller

Extract the first valid JSON object from LLM output. Handles: raw JSON, markdown fenced blocks, embedded in prose.

Number of trusted keys loaded.

List all conversations (most recent first).

Append an audit event. Returns `Ok(())` for a disabled logger (no path) and surfaces the failing IO step otherwise, so a security-critical caller can

Decide whether a proposed memory may be written by voice/tool flow.

Calculate RMS of a frame.

Check if a similar memory already exists (for deduplication). Uses SQL LIKE with key words from the content. More reliable than FTS5 for deduplicatio

Get all loaded skills (immutable).

Unique temp dir per test — avoids readonly SQLite flakes from fixed paths (#487).

Store an evergreen memory (never decays).

Transcribe a WAV file (works in all modes).