Summary
The personal-assistant profile has 4 layers: identity / long-term / short-term / ephemeral. Missing: a procedural tier for how-to knowledge and recurring routines.
Problem
Certain personal memories don't fit existing tiers:
| Memory |
Why Current Tiers Don't Fit |
| "WiFi password at the cabin is X" |
Not identity, not a preference (long-term), shouldn't decay in 7d (short-term) |
| "My gym routine: chest on Mon, legs on Wed" |
Procedural knowledge, needs ~30d half-life |
| "To submit expense reports, go to portal.company.com" |
How-to, not a preference |
| "The garage code is 4521" |
Factual reference, should persist longer than short-term |
Proposed Solution
Add a procedural layer between long-term and short-term:
layers:
identity:
half_life_days: 365
decay_model: power_law
long-term:
half_life_days: 90
decay_model: exponential
procedural: # NEW
half_life_days: 30
decay_model: exponential
description: "Routines, how-to knowledge, recurring workflows, reference codes"
short-term:
half_life_days: 7
decay_model: exponential
ephemeral:
half_life_days: 1
decay_model: exponential
Impact
- Extraction patterns (see related issue) should map routine/how-to patterns to this tier
- No breaking change — existing memories stay in their current tiers
- New tier only affects new saves with explicit
tier=procedural or auto-inferred
Effort
Small — YAML config change + update extraction tier inference + tests.
🤖 Generated with Claude Code
Summary
The
personal-assistantprofile has 4 layers: identity / long-term / short-term / ephemeral. Missing: aproceduraltier for how-to knowledge and recurring routines.Problem
Certain personal memories don't fit existing tiers:
Proposed Solution
Add a
procedurallayer betweenlong-termandshort-term:Impact
tier=proceduralor auto-inferredEffort
Small — YAML config change + update extraction tier inference + tests.
🤖 Generated with Claude Code