How do agent memory systems decide what to forget?

Short answer

They do not all forget the same way. Rank decay boosts recently used memories and dampens idle ones without removing rows. Expiration hides a memory after a date you set. Delete is an explicit remove. Invalidation marks a fact as no longer true while keeping the history that it once was. A context window that drops old tokens is not a forget policy; it is a prompt budget.

Four forget policies

Name the policy before you pick a product. The same word "forget" covers unlike operations.

• Decay: a search-time score. Mem0 Memory Decay is opt-in v3 search ranking: recently accessed memories are boosted (up to about 1.5x) and idle ones are dampened (floor about 0.3x). Decay never deletes or filters a candidate out. Docs: https://docs.mem0.ai/platform/features/memory-decay
• Expiration: hide after a date. Mem0 Memory Expiration hides a memory after a date you set; the row stays stored. Docs: https://docs.mem0.ai/platform/features/memory-expiration
• Delete: remove the row. Letta developers can delete archival passages via the SDK; agents cannot easily modify or delete archival memory. Docs: https://docs.letta.com/v1-sdk/memory/archival-memory/
• Invalidate: close a time window. Zep facts are edges with valid_at and invalid_at, so a later preference can supersede an earlier one without erasing history. Deleting an episode removes a derived fact only when no remaining episode supports it. Docs: https://help.getzep.com/facts

What is not a forget policy

Dropping tokens from the prompt because the window filled is not memory policy. It is a runtime limit. A JSON file the app rewrites by hand is not a product. Zep does not expire context on a default time schedule; retention policies and explicit deletes do. Docs: https://help.getzep.com/context-retention.md

How to choose

Match the policy to whether history must remain auditable.

• Need recency without loss: decay.
• Need a hide-after date: expiration.
• Need the row gone: delete.
• Need "this was true until Tuesday": invalidation with provenance.

Last reviewed

Facts on this page were checked against vendor documentation on 2026-10-28. Product surfaces move; treat version names and APIs as directional and follow the official links in Related.

Forget policies in agent memory (last reviewed: Oct 2026)

Where Pixeltable fits

Use Pixeltable when forget should be an ordinary table operation you own: insert memories, retrieve with EmbeddingIndex and a similarity threshold, delete or filter rows with where(), and keep transcripts beside them. Pixeltable is not Mem0, Letta, or Zep. It does not ship Memory Decay, sleep-time agents, or a temporal fact graph. Keep those products when that forget policy is the memory layer; store media or tool results in tables if you need them. Product comparison: Mem0 vs Letta vs Zep. Docs: Build an agent with memory.

import pixeltable as pxt
from pixeltable.functions.openai import embeddings, chat_completions

TableModel = pxt.model_base()

class Memories(TableModel, name='memories'):
    content: pxt.String
    category: pxt.String
    __indexes__ = [
        pxt.EmbeddingIndex(
            content,
            string_embed=embeddings.using(model='text-embedding-3-small'),
        ),
    ]

@pxt.query
def recall_memories(context: str, top_k: int = 3):
    sim = Memories.content.similarity(string=context)
    return (
        Memories.where(sim > 0.5)
        .order_by(sim, asc=False)
        .limit(top_k)
        .select(Memories.content, Memories.category)
    )

@pxt.udf
def build_prompt(memories: list[dict], user_message: str) -> str:
    memory_lines = '\n'.join(row['content'] for row in memories)
    return f'Relevant memories:\n{memory_lines}\n\nUser: {user_message}'

class Conversations(TableModel, name='conversations'):
    user_message: pxt.String
    relevant_memories = recall_memories(user_message, top_k=3)
    prompt = build_prompt(relevant_memories, user_message)
    response = chat_completions(
        messages=[{'role': 'user', 'content': prompt}],
        model='gpt-4o-mini',
    )
    assistant_reply = response.choices[0].message.content

Questions