Personal Memory Identity Resolution / 个人记忆身份解析

Definition

Personal memory identity resolution is the process of determining which person, relationship, and time a captured fact concerns before storing or recalling it as part of a personal AI’s memory.

Current Synthesis

Personal-agent inputs routinely mention multiple people. A health question may concern the user, a child, or a parent; an email or recording can mix the user’s preferences with a colleague’s statements; an old fact may be true of the same person only at an earlier time. Treating every sentence as a permanent fact about the account owner produces a coherent-looking but false profile.

The Today episode therefore frames memory as structured interpretation rather than accumulation. Identity and time need to travel with the fact, uncertainty should remain visible, and the system must be able to correct, supersede, decay, or delete earlier inferences. Graph-like relationship structure can help represent people, while event sequences need temporal structure; neither storage model eliminates the need for query-time judgment.

The EP53 failure case makes the identity risk concrete. An unstructured personal knowledge graph confused and duplicated people who shared a name. The rebuild represents people, companies, and tasks explicitly, then uses surrounding context to match incoming material against existing entities before creating or enriching a record. Identity resolution is therefore both a schema problem and an ingestion decision.

Key Claims

• Personal context often contains facts about people other than the account owner.
• Memory must distinguish event time from capture time and current state from historical state.
• Relationship facts and event sequences may need different representations.
• Uncertain attribution should not silently become a durable user-profile claim.
• Recall must select the right person’s information for the current query, not merely retrieve similar text.
• Inspection, correction, deletion, and supersession are necessary recovery mechanisms.
• New records should be matched against typed existing entities before creation, with contextual evidence used to avoid both mistaken merges and duplicates.

Evidence

• Person attribution: 277.从提醒你，到替你办：Today想把Personal AI带到哪一步？ uses health questions about the user, a child, or parents to show why all remembered information cannot be assigned to one person.
• Temporal attribution: 277.从提醒你，到替你办：Today想把Personal AI带到哪一步？ distinguishes when an event happened from when the system recorded it and argues that old preferences must yield to changed state.
• Representation and recall: 277.从提醒你，到替你办：Today想把Personal AI带到哪一步？ contrasts relationship graphs with event timelines and discusses dynamic context construction versus model-selected memory infrastructure.
• User recovery: 277.从提醒你，到替你办：Today想把Personal AI带到哪一步？ reports that Today exposes recorded memory for deletion and allows all memory to be cleared on exit.
• Graph ingestion failure: EP 53: Cognitive Surrender: The AI Engineering Mistake That Kills Projects describes same-name people being confused or duplicated when information entered a personal knowledge graph without explicit structure.
• Typed matching: EP 53: Cognitive Surrender: The AI Engineering Mistake That Kills Projects describes a rebuild that models people, companies, and tasks and uses surrounding context before creating or enriching an entity.

Counterevidence & Qualifications

The sources identify the problem but do not establish a solved architecture or evaluation benchmark. Graph, timeline, and query-time approaches create latency, compute, cache, and error tradeoffs. Contextual matching can still over-merge two people or fragment one person across nodes. User inspection can mitigate mistaken attribution but cannot replace robust defaults, especially when sensitive health, family, workplace, or bystander information is involved.

What Changed

• Added the same-name knowledge-graph failure as a concrete identity-resolution case.
• Added typed entities and context-based matching before record creation or enrichment.

Related Concepts

• Personal AI Memory - broader memory system whose facts require identity and temporal attribution.
• Personal Agent Understanding Layer / 个人Agent理解层 - interpretation layer that resolves and updates personal evidence.
• Context Decay - stale-context failure that identity and time metadata help expose.
• Identity Resolution Error - broader failure class for merging distinct real-world identities.
• Agent Permission Boundaries - limits capture and use of sensitive family, health, and workplace information.
• Consent-Based Recording - bystander-consent requirement when memory inputs include other people’s speech.
• AI Memory Lifecycle - maintenance layer that must preserve corrected identity and temporal state over time.