Pluggable memory engine for Elixir applications.
Provides document ingestion with markdown-aware chunking, configurable vector embeddings (pgvector), LLM-powered entity/relation extraction, knowledge graph storage, hybrid search, pipeline tracking, and memory decay.
Two Tiers
Tier 1 — Full Pipeline: Collections → Documents → Chunks → Entities → Relations. For structured ingestion of documents with chunking, embedding, entity extraction, and graph queries.
Tier 2 — Lightweight Knowledge: Entries + Edges. For simple knowledge storage with embeddings, access tracking, and edge traversal.
Quick Start
# Configure in your app
config :recollect,
repo: MyApp.Repo,
embedding: [provider: Recollect.Embedding.OpenRouter, model: "google/text-embedding-004"],
extraction: [llm_fn: &MyApp.LLM.chat/2]
# Lightweight API (Tier 2)
{:ok, entry} = Recollect.remember("The deploy script lives at scripts/deploy.sh",
scope_id: workspace_id, owner_id: user_id, entry_type: "note")
{:ok, results} = Recollect.search("deploy script", scope_id: workspace_id)
# Full Pipeline API (Tier 1)
{:ok, doc} = Recollect.ingest("Meeting Notes", long_markdown, owner_id: user_id)
{:ok, _run} = Recollect.process(doc)
{:ok, context} = Recollect.search("project timeline", owner_id: user_id)Summary
Functions
Format search results as text suitable for LLM system prompts.
Answer a question using memory search as LLM context.
Create an edge between two entries.
Archive stale entries based on access patterns.
Delete a knowledge entry.
Get the most recent session handoff.
Create a session handoff for continuing work across sessions.
Ingest content as a document with deduplication.
Run active invalidation to detect breaking changes and weaken related memories.
Run the learning pipeline to automatically extract knowledge from external sources.
Run the full pipeline on a document: chunk → embed → extract → sync.
Same as process/2 but runs asynchronously via TaskSupervisor.
Re-embed all entries/chunks with nil embeddings.
Store a knowledge entry with auto-embedding.
Hybrid search combining vector similarity and graph traversal.
Graph neighborhood expansion from an entity or entry.
Vector-only search across entries and/or chunks.
Functions
Format search results as text suitable for LLM system prompts.
Answer a question using memory search as LLM context.
Requires an :llm_fn callback that accepts a list of message maps
and returns {:ok, answer_string} or {:error, reason}.
Options
:llm_fn(required) —fn messages -> {:ok, string} | {:error, reason} end:system_prompt— Override the default system prompt:owner_id,:scope_id— Scope the search:limit— Max chunks (default: 10):hops— Graph depth (default: 2)
Create an edge between two entries.
Options
:weight— Float 0.0-1.0 (default: 1.0):metadata— Map of extra data
Archive stale entries based on access patterns.
Default: entries not accessed in 90 days with < 3 accesses.
Options
:max_age_days— Days since last access (default: 90):min_access_count— Minimum accesses to survive (default: 3)
Delete a knowledge entry.
Get the most recent session handoff.
Create a session handoff for continuing work across sessions.
Options
:scope_id- Required. The scope:what- What you were working on:next- List of next steps:artifacts- Files/links to continue with:blockers- What's blocking progress
Example
Recollect.handoff(workspace_id,
what: "Implementing user auth",
next: ["Add login controller", "Create session middleware"]
)Ingest content as a document with deduplication.
Returns {:ok, document} or {:ok, :unchanged} if content hash matches.
Options
:owner_id(required) — UUID of the owning user/workspace:collection_name— Collection name (default: "default"):source_type— "artifact", "conversation", "manual" (default: "manual"):source_id— External ID for dedup tracking
Run active invalidation to detect breaking changes and weaken related memories.
Options
:scope_id- Required. The scope to invalidate memories for:days- Number of days to scan (default: 7)
Example
{:ok, result} = Recollect.invalidate(scope_id: workspace_id)Run the learning pipeline to automatically extract knowledge from external sources.
Options
:scope_id- Required. The scope to learn for:sources- List of sources (default: all enabled):since- Learn from events since (default: "7 days ago"):dry_run- Preview without creating entries
Example
{:ok, result} = Recollect.learn(scope_id: workspace_id)Run the full pipeline on a document: chunk → embed → extract → sync.
Returns {:ok, pipeline_run} with tracking info.
Same as process/2 but runs asynchronously via TaskSupervisor.
Re-embed all entries/chunks with nil embeddings.
Store a knowledge entry with auto-embedding.
Options
:scope_id— UUID scope (workspace, user, etc.):owner_id— UUID of the owner:entry_type— String type (default: "note"):summary— Brief summary:source— "agent", "system", "user" (default: "system"):metadata— Map of extra data:confidence— Float 0.0-1.0 (default: 1.0)
Hybrid search combining vector similarity and graph traversal.
Searches both chunks (Tier 1) and entries (Tier 2) if available.
Options
:scope_id— Scope UUID (searches entries):owner_id— Owner UUID (searches chunks + entities):limit— Max results (default: 10):min_score— Minimum similarity score (default: 0.0):hops— Graph expansion depth (default: 1):tier—:full,:lightweight, or:both(default::both)
Graph neighborhood expansion from an entity or entry.
Vector-only search across entries and/or chunks.