Retrieval Boosting

Spaced repetition for AI

In human memory, recalling a fact strengthens it — making it harder to forget in the future. Cognitive Memory implements this through two-tier retrieval boosting.

Every time a memory is retrieved (appears in search results), its stability increases. The size of the boost depends on how long it’s been since the memory was last accessed.

Two boost tiers

Direct boost

Applied to memories that matched the search query directly (via vector similarity):

stability += 0.1 * spaced_rep_factor

Associative boost

Applied to memories that were activated through association links (not direct matches):

stability += 0.03 * spaced_rep_factor

The associative boost is intentionally smaller. These memories weren’t directly relevant to the query — they were pulled in because they’re linked to something relevant. They deserve some reinforcement, but less than a direct match.

Spaced repetition factor

The spaced_rep_factor rewards spaced practice over massed practice:

spaced_rep_factor = min(2.0, days_since_last_access / 7.0)

Days since last access	Factor
0 (same day)	0.0
1	0.14
3	0.43
7	1.0
14+	2.0 (capped)

This means:

• Retrieving a memory the same day gives almost no boost (you just saw it)
• Retrieving after 7 days gives the full base boost
• Retrieving after 14+ days gives 2x the base boost

This mirrors spaced repetition research: longer gaps between reviews produce stronger long-term retention.

What boosting affects

When a memory receives a boost, several things change:

1. stability increases (by the boost amount * factor)
2. access_count increments by 1
3. last_accessed_at updates to the current time
4. session_ids adds the current session ID

All of these feed back into the system:

• Higher stability slows future decay (see decay model)
• Higher access count contributes to core promotion eligibility
• Updated last_accessed_at resets the decay clock
• More session IDs move the memory toward the 3-session core promotion threshold

Worked example

A semantic memory with:

• stability = 0.3
• last accessed 10 days ago

Gets retrieved directly:

spaced_rep = min(2.0, 10/7) = 1.43
boost = 0.1 * 1.43 = 0.143
new_stability = min(1.0, 0.3 + 0.143) = 0.443

Gets retrieved associatively:

boost = 0.03 * 1.43 = 0.043
new_stability = min(1.0, 0.3 + 0.043) = 0.343

Over many retrievals, stability asymptotically approaches 1.0, making the memory increasingly resistant to decay.

Configuration

config = CognitiveMemoryConfig(
    direct_boost=0.1,              # base direct boost
    associative_boost=0.03,        # base associative boost
    max_spaced_rep_multiplier=2.0, # cap on spaced rep factor
    spaced_rep_interval_days=7.0,  # days for 1x factor
)

Cold reactivation

When a cold memory (in cold storage, not vector-searchable) is retrieved through an association link, it receives an associative boost and is migrated back to hot storage. This allows “forgotten” memories to be resurrected when contextually relevant.