Sleep Consolidation

Mnemosyne's background process that promotes aged Working Memory entries to Episodic Memory, inspired by human sleep's role in memory consolidation.

What Sleep Does

During each consolidation cycle, Mnemosyne:

1. Scans Working Memory for entries older than TTL/2 (12 hours by default)
2. Groups candidates by source
3. Summarizes each group via local LLM (TinyLlama) or AAAK text substitution
4. Promotes summaries to Episodic Memory (with summary_of tracking originals)
5. Evicts original Working Memory entries

Trigger Conditions

Consolidation triggers on:

• Age threshold: When Working Memory entries exceed TTL/2 (default: 12 hours with 24h TTL)
• Explicit call: mem.sleep() or mem.consolidate() for manual consolidation
• auto_sleep: Background thread that periodically runs consolidation without explicit calls (configure via config.yaml or MNEMOSYNE_AUTO_SLEEP env var)
• Shutdown: Graceful process exit

Consolidation Pipeline

flowchart LR
  A[Working Memory] -->|age > TTL/2| B[Candidate Selection]
  B -->|group by source| C[Group Entries]
  C -->|summarize| D[LLM or AAAK]
  D -->|promote| E[Episodic Memory]
  C -->|evict originals| F[Delete from Working]

Step-by-Step

1. Candidate selection: Find all Working Memory entries where created_at < now - TTL/2
2. Group by source: Cluster candidates by their source field
3. Summarize: For each group, generate a summary using:
   • Local LLM (TinyLlama) if available
   • AAAK encode() text substitution as fallback
4. Promote: Insert summary into episodic_memory with summary_of referencing the original IDs
5. Evict: Delete the original entries from working_memory
6. Log: Record the consolidation in consolidation_log

Retention Values

Monitoring Consolidation

Consolidation events are logged to the consolidation_log table: