Auto-Compaction

Prevent context rot by merging and discarding low-value memories.

As your agent runs, it accumulates memories. Some are high-value long-term facts, while others are transient noise or duplicates. Mneno’s Auto-Compaction engine helps keep your memory store lean and useful.

How Compaction Works

The compaction engine analyzes your memories and makes one of three decisions for each:

• Kept: The memory is high-value or recently updated and remains unchanged.
• Merged: Near-duplicate memories are combined into a single, consolidated memory.
• Discarded: Stale or low-scoring memories are removed to save space and reduce noise.

Preview vs. Compact

Mneno allows you to preview the changes before they are applied to your storage.

# Analyze memories and see what would happen
diff = client.preview_compaction()

print(diff.summary)
# Output: Compacted 100 memories into 60: kept 40, merged 30, discarded 30, created 10.

for decision in diff.discarded:
    print(f"Discarded {decision.memory_id}: {decision.reason}")

To actually apply the changes, use the compact() method:

diff = client.compact()
print(f"Reduction ratio: {diff.stats.estimated_reduction_ratio:.2%}")

Compaction Policy

You can control how aggressive compaction is by passing a CompactionPolicy.

from mneno.compaction import CompactionPolicy

policy = CompactionPolicy(
    max_memories=500,           # Hard limit on total memories
    min_score_to_keep=0.3,      # Discard anything scoring below this
    merge_duplicates=True,      # Consolidate similar memories
    stale_after_days=90,        # Discard low-importance memories older than this
    preserve_tags=["pinned"]    # Never discard memories with these tags
)

client.compact(policy=policy)

Policy Parameters

Parameter	Default	Description
max_memories	None	If set, forces the store down to this many memories, prioritizing high scores.
min_score_to_keep	0.20	Memories with a total score below this are candidates for deletion.
merge_duplicates	True	Whether to merge memories with high token overlap.
stale_after_days	180	Age at which low-importance memories are considered stale.
preserve_memory_types	[operational, preference]	Memory types that are never automatically discarded.
preserve_tags	[]	Tags that prevent a memory from being discarded.

Explainable Merges

When memories are merged, Mneno creates a new memory that consolidates the metadata and tags of the sources. The CompactionDiff includes a list of created memories and links them to the merged sources via related_memory_ids.