Memory Forget refers to the set of strategies and mechanisms that proactively remove, archive, or decay outdated, irrelevant, or low-value memories from an AI agent’s memory system. It is the deliberate process of making an AI “forget” things that no longer matter.
CN: Memory Forget（记忆遗忘）是指主动从 AI Agent 的记忆系统中移除、归档或衰减那些过时、无关或低价值记忆的一套策略和机制。它是让 AI 有意识地”遗忘”不再重要的事物的过程。
A well-designed Memory Forget system operates across three dimensions:
1. Temporal Decay (时间衰减)
Memories that haven’t been accessed for a certain period are marked as stale and candidates for removal. The longer a memory goes unused, the more likely it is to be forgotten.
CN: 超过一定时间未被访问的记忆被标记为过期（stale），成为清理候选。记忆越久未被使用，被遗忘的可能性越大。
Half-life — different types of memories have different decay rates:
| Category | Decay Rate λ | Half-life | Use Case |
|---|---|---|---|
| Strategy | 0.10 | ~38 days | Proven methodologies, architecture patterns |
| Fact | 0.16 | ~24 days | User preferences, identity, tech choices |
| Assumption | 0.20 | ~19 days | Inferred context, uncertain info |
| Failure | 0.35 | ~11 days | Lessons learned, environment issues |
| 类别 | 衰减率 λ | 半衰期 | 适用场景 |
|---|---|---|---|
| Strategy（策略） | 0.10 | ~38天 | 被验证有效的方法论、架构模式 |
| Fact（事实） | 0.16 | ~24天 | 用户偏好、身份信息、技术选型 |
| Assumption（推断） | 0.20 | ~19天 | 推断的上下文、不确定的信息 |
| Failure（失败教训） | 0.35 | ~11天 | 踩过的坑、环境问题 |
2. Importance Scoring (重要性评分)
Not all memories are equal. Each memory gets a score based on:
- Source: Explicit (user directly stated it) > Inferred (AI deduced it)
- Confidence: How certain is this memory?
- Access Frequency: How often is this memory recalled?
CN: 不是所有记忆都同等重要。每条记忆基于以下因素获得评分：
- 来源：显式（用户直接说的）> 推断（AI 推测的）
- 置信度：这条记忆有多确定？
- 访问频率：这条记忆被召回的频率有多高？
EN: When capacity limits are reached, memories with the lowest composite scores are cleaned first.
3. Capacity Limitation (容量限制)
Each user or session has a maximum memory limit. When exceeded, the system evicts the lowest-value memories to make room for new ones.
CN: 每个用户或会话有最大记忆上限。超出时，系统驱逐价值最低的记忆，为新记忆腾出空间。
Forgetting mechanism design patterns
遗忘的实现模式
There are two complementary loops for keeping the memory system healthy:
1. 自动循环（Automatic Loop）
EN: Background processes that run periodically (e.g., daily at 3 AM) to:
- Check decay scores for all memories
- Re-classify memories (is it still true?)
- Prune the memory graph (remove dead links)
- Archive or delete decayed memories
- 检查所有记忆的衰减分数
- 重新分类记忆（它仍然是真的吗？）
- 修剪记忆图（移除死链接）
- 归档或删除已衰减的记忆
2. 人工循环（Human Loop）
EN: Weekly reviews where a human reviews uncertain memories in an “inbox” to verify or delete them.
CN: 每周一次的人工审查，人类审阅”收件箱”中不确定的记忆，验证或删除它们。