L1: Memory Evolution
Layer 1 is the most frequent and lowest-risk self-evolution layer. It operates on the agent's memory system, automatically compacting redundant entries and clustering related memories by topic.
Overview
L1 evolution runs after every session (or on a configurable schedule) and performs:
- Compaction -- merge multiple related memory entries into concise summaries
- Topic clustering -- group memories by semantic similarity
- Relevance scoring -- adjust memory weights based on access frequency
- Pruning -- remove memories that have become stale or contradicted
How It Works
- After a session ends, L1 analyzes the newly stored memories
- It identifies clusters of related entries using embedding similarity
- Clusters that exceed a size threshold are compacted into summaries
- Memory relevance scores are updated based on recall frequency
Configuration
toml
[self_evolution.l1]
enabled = true
schedule = "after_session" # or "hourly", "daily"
compaction_threshold = 10
cluster_similarity = 0.8
min_access_count = 2