Memory Core

What this surface is for

Background service

Behavior in the current codebase

• Incrementally mines durable facts from prior conversations instead of relying only on the active turn.
• Builds user and room summaries in Redis for later Hydra injection.
• Includes confidence thresholds, identity linking options, and context-size limits.

Direct core support

Memory Core is background infrastructure. It scans chat history, extracts durable facts, and injects memory context back into Hydra instead of acting like a direct Verba surface.

Configuration schema

• Auto-link Identity by Name checkbox

  Automatically map matching usernames across portals to one shared identity profile.

  Key: auto_link_identities Default: False
• Extraction Max Tokens number

  Max completion tokens for memory extraction.

  Key: extraction_max_tokens Default: 2700
• Hydra Memory Items number

  Max memory facts injected per scope (user and room).

  Key: hydra_max_items Default: 12
• Hydra Summary Chars number

  Max characters for user/room memory summaries in context.

  Key: hydra_summary_max_chars Default: 2100
• Hydra Value Chars number

  Max characters per memory fact value in context.

  Key: hydra_value_max_chars Default: 288
• Interval (sec) select

  How often memory extraction runs.

  Key: interval_seconds Default: 180 Options: 1 minute, 2 minutes, 3 minutes, 5 minutes, 10 minutes, 15 minutes, 30 minutes, 1 hour
• Lookback Limit number

  Max new messages processed per scope each pass.

  Key: lookback_limit Default: 80
• Min Confidence number

  Minimum confidence required before storing facts.

  Key: min_confidence Default: 0.65
• Write Room Memory checkbox

  Enable writes to mem:room:{platform}:{room_id}.

  Key: write_room_memory Default: True
• Write User Memory checkbox

  Enable writes to durable user docs. With auto-link on: mem:user:identity:{identity_id}; with auto-link off: mem:user:{platform}:{user_id}.

  Key: write_user_memory Default: True