Dynamic Forgetting

Hebb Mind implements a dynamic forgetting mechanism inspired by the Ebbinghaus forgetting curve. Memories that are rarely accessed and low in importance gradually decay and are removed, while frequently accessed, high-importance memories persist.

The Formula

Each memory's time-to-live is computed dynamically:

TTL = base_ttl * (1 + log(access_count)) * (importance / 5) * exp(-decay_factor * days_since_access)

Where:

• base_ttl -- base time-to-live in hours (default: 168 hours / 7 days)
• access_count -- number of times the memory has been accessed (retrieved in search results)
• importance -- LLM-rated score from 0 to 10
• decay_factor -- exponential decay rate (default: 0.693, corresponding to a half-life of 1 day)
• days_since_access -- days since the memory was last accessed

How It Works

The forgetting job runs periodically and evaluates every memory:

1. Computes the dynamic TTL using the formula above
2. Compares the TTL against the time elapsed since the memory was last accessed
3. Removes memories whose TTL has expired

Factors That Extend a Memory's Life

• High access count -- each retrieval increases the log(access_count) term, reinforcing the memory
• High importance score -- memories rated 8/10 live 60% longer than those rated 5/10
• Recent access -- the exponential decay term exp(-decay * days) is largest right after access

Factors That Shorten a Memory's Life

• Never accessed -- log(1) = 0, so the access bonus is zero
• Low importance -- a score of 2/10 gives only 40% of the base TTL
• Long time since last access -- the exponential decay rapidly reduces effective TTL

Configuration

Field	Default	Description
base_ttl_hours	168	Base TTL in hours (7 days)
decay_factor	0.693	Exponential decay factor
forget_interval_seconds	1800	How often the forgetting job runs (30 minutes)

# Memories live longer (14 days base)
hebb config set base_ttl_hours 336

# Slower decay (memories fade more gradually)
hebb config set decay_factor 0.3

# Run forgetting less frequently (every hour)
hebb config set forget_interval_seconds 3600

Manual Trigger

Trigger the forgetting job immediately:

curl -X POST http://localhost:8321/api/v1/admin/forget

Example Scenarios

Scenario 1: Frequently accessed, important memory

A memory with importance 8, accessed 10 times, last accessed 2 days ago:

TTL = 168 * (1 + log(10)) * (8/5) * exp(-0.693 * 2)
    = 168 * 3.30 * 1.6 * 0.25
    = 221.8 hours (~9.2 days remaining)

This memory is still well within its TTL.

Scenario 2: Neglected, low-importance memory

A memory with importance 3, accessed once, last accessed 5 days ago:

TTL = 168 * (1 + log(1)) * (3/5) * exp(-0.693 * 5)
    = 168 * 1.0 * 0.6 * 0.031
    = 3.1 hours

This memory has effectively expired and will be removed.

Design Rationale

Traditional memory systems either keep everything forever or require manual cleanup. Hebb Mind's dynamic forgetting provides:

• Automatic cleanup -- no manual memory management needed
• Adaptive retention -- important, frequently-used memories survive naturally
• Bounded storage -- database size stays manageable over time
• Biological plausibility -- mirrors how human memory works