Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.linksee.app/llms.txt

Use this file to discover all available pages before exploring further.

Sleep-mode compression. Clusters cold, low-importance memories by (entity, layer), summarizes each cluster into a single protected learning-layer entry, deletes originals, and runs a forget-sweep. Run at session end or on demand. Set dry_run: true to preview without writing.

​
Parameters

​
scope
string
default:"session"
all — consolidate across all entities. session — only the current session’s entities.
​
min_age_days
number
default:"7"
Minimum age for memories to be eligible for clustering. Set to 0 to consolidate immediately (useful after bulk import).
​
dry_run
boolean
default:"false"
Preview what would be compressed without modifying the database.
​
use_llm
boolean
default:"false"
If true, uses MCP Sampling to request the client LLM to write better consolidated summaries. Falls back to heuristic if the client refuses.

​
What gets consolidated

LayerEligible?Notes
contextYesSituational context fades over time
emotionYesEmotional state is session-specific
implementationYesHow-details get compressed to learnings
goalNoGoals are never consolidated
caveatNoCaveats are permanently protected
learningNoAlready the target layer

​
Eligibility criteria

  • Heat score < 30 (cold)
  • At least 2 memories in the cluster
  • Age >= min_age_days

​
Output

Each consolidated cluster produces a learning-layer memory containing:
  • Exemplar fragments from the original memories
  • Time period covered
  • Count of originals replaced
  • replaced_ids for audit trail

​
Post-consolidation

After clustering, consolidate also:
  1. Forget-sweep: Runs auto-forgetting on remaining unprotected memories
  2. Stalled detection: Marks in_progress memories untouched for 30+ days as stalled

​
Example

{
  "scope": "all",
  "min_age_days": 14,
  "dry_run": true
}
Run consolidate weekly or use the weekly-consolidation prompt for a guided workflow.