Decay, Pruning, and Forgetting on Purpose

Memory grows without bound; relevance does not

An always-on AI accumulates thousands of memory items per week. Most are stale within a month. Retrieval quality degrades as the irrelevant pile grows — the embedding model is good but not magical.

Three pruning strategies

Hard age cutoff — delete anything older than N days unless flagged as keep-forever. Brutal, simple, recoverable from JSONL ground truth.
Distillation + replace — periodically compress old episodes into distilled facts and delete the raw turns. Permanent loss of detail in exchange for cleaner retrieval.
Use-based decay — track how often each memory was retrieved + whether the LLM cited it; prune cold items. Best retention quality; needs retrieval logging.

The keep-forever escape hatch

Some memories must never decay: identity ('user is Dad', 'I am Pippa'), policies ('never auto-push to main'), past traumas/lessons. Mark these with metadata.permanent = true and exclude them from any pruning pass. Without this, your AI quietly forgets who it is.

Code

Pruning pass with permanent escape·python

from datetime import datetime, timezone, timedelta

def prune(collection, max_age_days: int = 90):
    cutoff = (datetime.now(timezone.utc) - timedelta(days=max_age_days)).isoformat()
    candidates = collection.get(
        where={
            '$and': [
                {'created_at': {'$lt': cutoff}},
                {'permanent': {'$ne': True}},
            ],
        }
    )
    if candidates['ids']:
        collection.delete(ids=candidates['ids'])
        print(f'pruned {len(candidates["ids"])} stale memories')

Decay, Pruning, and Forgetting on Purpose

Memory grows without bound; relevance does not

Three pruning strategies

The keep-forever escape hatch

Code

External links

Exercise

Progress

Comments 0