​
What exactly is a “memory”?

A memory is the smallest unit of information you store in Flumes—typically a chat message, document chunk, or JSON blob. Each memory is stored with:
  • id – deterministic UUID
  • memory – raw text content used for embedding & display
  • metadata – arbitrary JSON for filtering (e.g. {"source":"chat"})
  • user_id, agent_id, run_id – scoping keys that isolate data across tenants/agents
  • Timestamps: created_at, updated_at

​
Embeddings & similarity

When you create a memory we automatically:
  1. Clean & truncate the text to fit the model context window.
  2. Generate an embedding using an in-house model (compatible with OpenAI text-embedding-3-small).
  3. Upsert the vector into a FAISS index (
    Postgres + pgvector coming soon
    ).
During retrieval we perform cosine similarity search with optional metadata filters and time decay.

​
Hot vs cold memory

Cold storage is on the roadmap. For now all memories live in the “hot” (immediate) index.
TierLatencyUse-case
Hot< 200 msChat history, agent state
Cold~1 sLong-term knowledge base

​
Decay & summarization (coming soon)

We’re adding built-in time decay and auto-summarization so that stale memories roll up into compact summaries, keeping your index lean and latency snappy.
Follow our changelog for upcoming memory operators.