Skip to main content

The Memory Working Set

The Working Set is the finalized, ranked collection of memories that are injected into an AI agent's prompt context for a specific conversation or task.

It represents the system's "best guess" at the knowledge the AI needs right now to be helpful and accurate.

Definition​

A working set is dynamic and ephemeral. It is generated on-the-fly by the Hybrid Retrieval system based on:

  • The current user and organization (The Subject).
  • The current conversation history or query.
  • The available memory blocks for that tenant.

Characteristics​

  • Ranked: Memories are ordered by their composite score (Similarity + Salience + Confidence).
  • Capped: The working set has a limit (e.g., 10-20 blocks) to prevent context window saturation and "lost in the middle" effects.
  • De-aliased: The system ensures only the latest version of a memory is present if multiple SUPERSEDES relationships exist.
  • Aggregated: Memories may be grouped into Memory Packs for better structural presentation to the AI.

The Generation Lifecycle​

  1. Trigger: An agent calls the memory.query tool.
  2. Retrieve: The hybrid engine fetches structurally and semantically relevant blocks.
  3. Filter: Non-salient or low-confidence blocks are discarded.
  4. Rank: Remaining blocks are scored and sorted.
  5. Finalize: The top-N blocks are compiled into the final prompt context.

Working Set in the UI​

In the Memory Browser, users can see the "Current Working Set" for their session. This allows administrators to:

  • See exactly what context the AI is currently "remembering."
  • Inspect the scores for each block in the set.
  • Manually pin or remove blocks if the Working Set is inaccurate.

Best Practices for Developers​

  • Manage Context Size: Don't request too many blocks at once. A focused working set of 5-10 high-quality memories is often more effective than 50 noisy ones.
  • Use Subjects: Always query within a subject scope to ensure the working set is relevant to the current conversation.
  • Trust the Scoring: Let the engine handle the ranking and salience decay to ensure the most "important" information naturally stays in the set.