Modules
Pluggable memory engine for Elixir applications.
LLM-free memory classification using regex-based pattern matching.
Confidence lifecycle management.
Runtime configuration resolution for Recollect.
Automatic conflict detection for contradictory memories.
Conflict management for contradictory memories.
Sleep consolidation for memory maintenance.
Detect current environment context from the running system.
Behaviour for database-specific implementations.
libSQL/SQLite database adapter with native vector support.
PostgreSQL database adapter with pgvector support.
SQLite database adapter with sqlite-vec extension for vector support.
Local embedding provider using Bumblebee with sentence-transformers/all-MiniLM-L6-v2.
OpenRouter embedding provider.
Behaviour for embedding providers.
Ecto type for embedding vectors that works with both PostgreSQL and libSQL.
Export Recollect data to a portable format for migration or backup.
Default extraction provider using LLM with JSON structured output.
Behaviour for LLM-powered entity/relation extraction.
Graph store implementation using PostgreSQL recursive CTEs. Queries the recollect_entities and recollect_relations tables directly.
Behaviour for graph storage backends.
Session handoff for continuing work across sessions.
Import Recollect data from JSONL export files.
Active memory invalidation based on detected or explicit breaking changes.
Lightweight knowledge API (Tier 2). Simple store-embed-search for entries and edges.
Behaviour for learning modules that extract knowledge from external sources.
Unified learner for all coding agents (Claude Code, Codex, Gemini CLI, Goose, OpenCode).
Provider for Claude Code (~/.claude/).
Provider for OpenAI Codex CLI (~/.codex/).
Provider for Google Gemini CLI (~/.gemini/).
Provider for OpenCode (~/.local/share/opencode/).
Behaviour for coding agent provider modules.
Learn from git history — commits, branches, tags.
Groups git commits into synthesized development insights.
Detects the project's technology stack from config files and tracks framework transitions across git history.
Orchestrate learning from multiple sources.
Archives stale entries based on access patterns. Entries not accessed in N days with fewer than M accesses are archived.
Re-embeds entries, chunks, and entities using a pluggable embedding
function and tracks per-row provenance via embedding_model_id.
Generates database-specific migration content for Recollect tables.
Context mipmaps — progressive detail levels for retrieval.
Outcome feedback for closing the learning loop.
ETS table for tracking the last-retrieved entry IDs per scope.
Orchestrates the full memory ingestion pipeline.
Splits document text into chunks suitable for embedding and retrieval.
Embeds chunks and entities using the configured embedding provider. Supports batch processing and async embedding via TaskSupervisor.
Extracts entities and relationships from text chunks using LLM structured output. Deduplicates and persists results with mention counting.
Ingests content into the memory system with content-hash deduplication. Creates or updates MemoryDocuments within a collection.
GenServer that owns an ETS :counter table for retrieval bumps.
Text fragment with vector embedding, created by the Chunker. Preserves section hierarchy and paragraph boundaries.
Groups related memory documents. One per user/workspace/topic.
Source content record. Tracks original artifacts/conversations being indexed. Content hash enables deduplication — unchanged documents skip re-processing.
Lightweight knowledge edge between entries (Tier 2).
Named entity extracted from chunks via LLM. 10 entity types with mention counting and prominence tracking.
Lightweight knowledge entry (Tier 2). Simple store-embed-search with access tracking and decay support.
Tracks pipeline execution: status, step details, cost, and timing.
Graph edge between entities. 8 typed relations with confidence weights.
Schema fit computation for new entries.
ETS table for schema acceleration data.
Unified search combining vector similarity, graph traversal, and edge following.
LLM-augmented retrieval that combines hybrid search with LLM reasoning.
Calculate context boost for search results based on entry hints matching current context.
Formats search results as text suitable for LLM system prompt injection.
Graph-based search using PostgreSQL recursive CTEs. Provides neighborhood expansion and relation queries.
Semantic similarity search over chunks and entries.
Memory strength calculation based on hippocampal-inspired mechanics.
Telemetry events emitted by Recollect.
Emotional valence inference and handling.
Working memory layer (Tier 0).
GenServer holding a bounded working memory buffer for a single scope.
Mix Tasks
Run Recollect sleep consolidation for a scope.
Generates Recollect database migrations in the host application.