Knowledge Erasure

Clear, delete, and inspect knowledge with loud, predictable guarantees

Erasure clears an agent’s knowledge base between sessions, with clear errors when an adapter can’t support an operation.

from praisonaiagents import Agent

agent = Agent(name="privacy-aware", instructions="Answer, then clear context between sessions.", knowledge=["policy.md"])
agent.start("Summarise the policy")
agent.knowledge.reset()   # clears the knowledge base between sessions

Quick Start

1. Reset the knowledge base

   from praisonaiagents import Agent
   
   agent = Agent(name="assistant", instructions="Answer from the docs", knowledge=["policy.md"])
   agent.start("Summarise the policy")
   
   # Clear the knowledge base between sessions
   agent.knowledge.reset()

2. Delete all scoped memories

   # Delete everything for a given scope
   agent.knowledge.delete_all(user_id="user-123")

3. Inspect change history

   # History is only available on Mem0-backed knowledge
   changes = agent.knowledge.history(memory_id="mem-abc")

How It Works

Knowledge erasure calls route to the underlying memory adapter. Unsupported operations fail loudly instead of raising AttributeError.

Adapter Capabilities

Support differs per adapter — verify against praisonaiagents/knowledge/adapters/ before relying on an operation.

Common Patterns

Guard erasure calls so your code works across adapters:

from praisonaiagents import Agent

agent = Agent(name="assistant", instructions="Answer from the docs", knowledge=["policy.md"])

# reset() never raises — it warns and no-ops on adapters without native reset
agent.knowledge.reset()

# delete_all() / history() may not be supported
try:
    agent.knowledge.delete_all(user_id="user-123")
except NotImplementedError:
    # Fall back to reset() on adapters without scoped delete
    agent.knowledge.reset()

Report to users honestly on a destructive path — check the return value before claiming success:

from praisonaiagents import Agent

agent = Agent(name="assistant", knowledge=["policy.md"])

# Never claim success on a destructive path without checking
if agent.knowledge.reset():
    print("Knowledge cleared")
else:
    backend = type(agent.knowledge.memory).__name__
    print(f"Knowledge was NOT cleared — backend {backend} has no reset()")

Best Practices

Related