Four typed memory layers beat one vector store (for personal AI assistants)

Description and use

Splitting memory into four typed layers—facts, events, lessons, and inferences—eliminates contradictions, stale events, missed corrections, and un-decaying inferences, resulting in a more accurate and reliable personal AI assistant over time.

Problem

A flat vector store for personal AI assistants mixes facts, events, lessons, and inferences, causing contradictory facts, stale events, missed corrections, and un-decaying inferences, which degrade assistant reliability over time.

Outcome

Typed memory layers eliminate contradictory facts, prevent stale events, ensure lessons override behavior, and decay unverified inferences, leading to a sharper, more reliable personal AI assistant over time.

Workflow

• 1. Store facts using upsert on (subject, predicate) to overwrite old values.
• 2. Append events with timestamps; retrieve via time‑window or explicit recall.
• 3. Append lessons; inject them into context whenever their importance exceeds a threshold.
• 4. Append inferences with confidence; run a decay job to reduce confidence over time.
• 5. Use separate read paths for each type to construct the assistant’s context.
• 6. Compact and prune old entries to keep the SQLite file bounded (~117 MB after 6 months).

Evidence

Four tables, four write paths, four read paths: Fact, Event, Lesson, Inference

Source

The upsert rule for facts is the single highest-leverage line in that table.

Source