Tokenize French/English text for BM25 search. Features: - Lowercases text - Normalizes accents for matching - Extracts words and

Load vectors from .npz file. Returns: True if loaded successfully, False if file doesn't exist or numpy unavailable

Extract named entities from chunk content (Phase 7.2, MAGMA-inspired). Uses regex patterns to identify domain-specific entities: files,

Check if a chunk falls within a date range. Uses simple string comparison on YYYY-MM-DD format (lexicographic = chronologic). Args:

Search chunks (and optionally insights) using BM25 ranking. Args: query: Natural language search query top_k

Check if a chunk's entities contain a given entity (Phase 7.2). Performs case-insensitive substring matching across all entity types. A

Archive a chunk by compressing it to archive/. Steps: 1. Read chunk file 2. Compress to archive/*.md.gz 3. Remove from main inde

Add a vector for a chunk. If chunk_id already exists, replaces the vector. Args: chunk_id: The chunk identifier

Save content to an external chunk file. Use this to externalize parts of the conversation that you might need to reference later but don

Write JSON atomically using write-to-temp-then-rename. On POSIX systems, rename is atomic, so the file is never in a half-written state.

Remove accents from text for matching purposes. Uses NFD normalization to decompose characters, then removes diacritical marks (combinin

Helper to create a chunk for testing.

Get or create the cached embedding provider (singleton, lazy). Reads RLM_EMBEDDING_PROVIDER env var (default: "model2vec"). Returns None if t

Embed a list of texts into vectors. Args: texts: List of text strings to embed Returns: numpy ndarray of sha

Load archive index from JSON file.

Extract date string (YYYY-MM-DD) from chunk metadata or ID. Tries created_at first, then falls back to parsing the chunk ID. Args:

Load chunks index from JSON file.

Load chunks index from JSON file.

Normalize BM25 scores to [0, 1] range using min-max scaling. Adds a 'score_norm' field to each result dict. Args: results: List of B

Return the embedding dimension.

Determine if a chunk is protected from archiving/purging. A chunk is IMMUNE if ANY of these apply: - Has a protected tag (critical, deci

Load session memory from JSON file.

Load sessions index from disk.

Restore an archived chunk back to active storage. Steps: 1. Find in archive index 2. Decompress to chunks/ 3. Remove from archiv

Execute retention actions. Args: archive: Archive old unused chunks (default: True) purge: Purge very old archives (default:

Persist vectors atomically to .npz file.

Validate chunk ID format to prevent path traversal. Allows: alphanumeric, hyphens, underscores, dots, ampersands. Blocks: slashes, "..",

Create an active chunk file + index entry old enough to archive.

Helper to create a chunk for testing.

Helper to create a chunk file and index entry.

Get list of chunks eligible for archiving. A chunk is an archive candidate if: - Age > ARCHIVE_AFTER_DAYS (30 days) - access_count =

List all available chunks with their metadata. Use this to see what conversation history is available in external storage. Args:

Get translated string for current language.

Helper to create a chunk.

Extract content from a chunk file, skipping YAML header. Phase 8.1: Prepends summary, tags, project, and domain from the YAM

Perform semantic vector search if available. Returns None if semantic search is not available (deps missing), allowing the caller to fall bac

Load purge log from JSON file.

Save archive index atomically.

Save sessions index atomically.

Build BM25 index from all chunks and optionally insights. Reads all .md files in chunks directory, tokenizes content, and bu

Read content from a chunk file. Use this to view the contents of a previously saved chunk. Can read the whole chunk or just a portion (b

Permanently delete an archived chunk. Steps: 1. Find in archive index 2. Log metadata to purge_log.json (NOT content) 3. Remove

Preview retention actions without executing. Shows what would be archived or purged based on current rules. Returns: Dictionary

Build a safe file path from chunk ID, validating against traversal. Args: base_dir: The directory the file must remain within

Helper to create a chunk.

Helper to create a chunk file and index entry.

Auto-detect project name from environment or git (Phase 5.5). Priority: 1. RLM_PROJECT environment variable (explicit override) 2. G

Create one chunk per date, all containing the term 'deployment', appended to the index in chronological (oldest-first) order.

Save chunks index atomically.

Save session memory atomically.

Drop a chunk's vector from the semantic store. Never raises.

Extract content from a chunk file, enriched with YAML metadata. Phase 8.1: Prepends summary and tags from YAML header to improve embedding qu

Remove an insight from memory. Args: insight_id: The ID of the insight to remove Returns: Confirmation of removal

Get list of archived chunks eligible for purging. A chunk is a purge candidate if: - In archive for > PURGE_AFTER_DAYS (180 days) -

Search for a pattern across all chunks. Use this to find where a topic was discussed or where specific information is stored. Phase

List all available domains from domains.json. Returns: dict with domains organized by category

List available sessions with optional filtering. Args: project: Filter by project name domain: Filter by domain limi

Entry point for the MCP RLM server.

Get current status of the memory system. Returns: Statistics about stored insights

Retrieve insights from memory. Use this to: - Find information discussed earlier in conversation - Review decisions made - Check

Save an important insight to persistent memory. Use this to remember: - Key decisions made during conversation - Important facts dis

Remove a chunk's vector. Args: chunk_id: The chunk identifier to remove Returns: True if found and removed,

Restore an archived chunk back to active storage. Args: chunk_id: ID of the archived chunk Returns: Dictionary with sta

Search for nearest vectors using cosine similarity. Args: query_vec: 1D numpy array (query embedding) top_k: Maximum

Generate automatic summary from content (Phase 4.1). Extracts the first non-empty line as summary. If too long, truncates with ellipsis.

Check if content with this hash already exists (Phase 4.2). Args: content_hash: MD5 hash of normalized content Returns:

Generate hash of normalized content for duplicate detection (Phase 4.2). Normalizes whitespace and lowercases before hashing to catch ne

Rough token estimation (1 token ~ 4 chars for French/English).

Extract summary from chunk YAML header. Args: chunk_file: Path to the chunk .md file Returns: Summa

Generate a unique chunk ID (Phase 5.5 enhanced). Format 2.0: {date}_{project}_{seq}[_{ticket}][_{domain}] Args: project: Projec

Generate a short unique ID for an insight.

Increment access counter for a chunk (Phase 4.3). Uses file locking to prevent race conditions on concurrent access. Args: chun

Load domains configuration from disk. Creates default if not exists.

Save chunks index atomically.

Save purge log atomically.

(Re)embed a restored chunk into the semantic store. Never raises. Mirrors the enrichment used at chunk creation time (tags + summary + body)

Add a chunk ID to a session's chunk list. Args: chunk_id: The chunk ID to add session_id: Target session (uses current if no

Write text file atomically using write-to-temp-then-rename. Args: filepath: Target file path content: Text content to write

Run full benchmark for one provider.

Compare search results between providers.

Extract content from a chunk file, skipping YAML header.

Read context usage from stdin (passed by Claude Code).

Fuzzy grep - find matches even with typos (Phase 5.2). Uses thefuzz library for approximate string matching. Tolerates typos like "valid

Check if the target file is in an auto-memory directory.

Load all chunks (id, content) from index.

Context manager for locked read-modify-write on a JSON file. Acquires an exclusive lock, yields the data for modification, then writes b

Entry point for the MCP RLM server.