Bounded Learning

Keep continuous-learning stores healthy by capping size and expiring stale entries.

Keep continuous-learning stores healthy by capping size and expiring stale entries — automatically, without any LLM judgement.

from praisonaiagents import Agent, LearnConfig

agent = Agent(
    name="Assistant",
    instructions="Adapts without unbounded memory growth",
    learn=LearnConfig(persona=True, max_entries=1000, retention_days=90),
)
agent.start("Remember I prefer bullet points")

The user keeps teaching the agent; old or excess learnings archive automatically so search stays fast.

How It Works

Quick Start

1. Enable with a cap

   Set max_entries to limit how many learnings each store retains:

   from praisonaiagents import Agent, LearnConfig
   
   agent = Agent(
       name="Assistant",
       instructions="Adapts to user preferences",
       learn=LearnConfig(
           persona=True,
           max_entries=1000,    # Keep at most 1000 entries per store
           retention_days=90,   # Mark entries older than 90 days as stale
       ),
   )
   
   agent.start("Help me plan my day")

2. Run a manual prune

   Call prune() on the learn manager for one-off cleanup:

   from praisonaiagents import Agent, LearnConfig
   
   agent = Agent(
       learn=LearnConfig(max_entries=1000, retention_days=90)
   )
   
   # Returns evicted count per store
   evicted = agent.memory.learn().prune()
   print(f"Pruned: {evicted}")
   # e.g. {"persona": 12, "insights": 3}

How It Works

Usage telemetry is observed automatically — every search() and list_all() call increments use_count and updates last_used on returned entries. No extra configuration needed.

Eviction scoring: entries are ranked by use_count ascending, then created_at ascending. The least-used, oldest entries are evicted first — the most valuable, recently accessed learnings always survive.

Archival: evicted entries are appended to a per-store .archive.json sidecar (e.g., persona.archive.json) in the same directory as the active store file. The archive is append-only JSON — nothing is permanently hard-deleted. Writes are atomic and locked per store, so a crash mid-write cannot destroy earlier archived entries and two processes sharing a store cannot race. If the existing archive is unreadable (torn JSON or not a list), it is moved aside to a .corrupt sidecar and eviction carries on — max_entries and retention_days keep being enforced. Added in PraisonAI PR #6172.

Configuration Options

Full reference on max_entries and retention_days is in the LearnConfig reference.

from praisonaiagents import LearnConfig

LearnConfig(
    max_entries=0,      # 0 = unbounded (default)
    retention_days=0,   # 0 = never expires (default)
)

Best Practices

learn=LearnConfig(
    persona=True,
    insights=True,
    max_entries=1000,
)

learn=LearnConfig(
    persona=True,
    max_entries=500,
    retention_days=30,  # 30-day data retention policy
)

~/.praisonai/learn/private/default/persona.archive.json

Related