Decay & Reinforcement

Control memory lifespan with decay rates and reinforcement

Decay & Reinforcement

What makes Memory Module different: Memories that matter stay strong through reinforcement, while noise fades away through natural decay.

Memories naturally fade over time (decay). Access them or manually reinforce them to keep them strong. Different memory types fade at different rates.

---

Understanding Salience

Salience = importance score (0.0-1.0)

Range	Status	Effect
1.0	Brand new	Maximum importance
0.7-0.9	Strong	Easily accessible
0.3-0.6	Medium	Still useful
0.1-0.2	Weak	Fading
Under 0.1	Candidate	Auto-pruning

Why it matters:

1. Whether memory appears in search
2. How it ranks when it does appear
3. When it gets auto-pruned

---

Natural Decay: Forgetting Curve

Every memory decays exponentially, just like human memory.

Sector	Decay Rate (λ)	Half-Life	Example
Reflective	0.001 (slowest)	~693 days (~2 years)	“Why we chose microservices”
Semantic	0.005	~139 days (~4.6 months)	“JWT tokens expire after 30 days”
Procedural	0.008	~87 days (~2.9 months)	“How to deploy to production”
Episodic	0.015	~46 days (~1.5 months)	“Sprint planning meeting”
Emotional	0.020 (fastest)	~35 days (~1.2 months)	“Customer frustrated”

Half-life: Time for salience to drop to 50%. Memory continues fading exponentially after half-life.

Decay Examples

Reflective (Slow)

Strategic decision:

Day 0:   Salience 1.00 ████████████████████
Month 3: Salience 0.92 ██████████████████
Month 6: Salience 0.85 █████████████████
Year 1:  Salience 0.72 ██████████████
Year 2:  Salience 0.52 ██████████

Still strong after 2 years!

---

Reinforcement: Strengthening Memories

Automatic Reinforcement

Happens every time you access a memory:

• Searching and finding it
• Retrieving by ID
• Traversing via waypoints
• Viewing in admin panel

Effect: Small salience boost (+0.02 to +0.05), timestamps update

Example:

Your team's API auth docs (Semantic):

Created:     Salience 1.00
Month 2:     Salience 0.95 (natural decay)
Accessed 3x: Salience 0.98 (reinforced through use)
Month 4:     Salience 0.94 (decay continues)
Accessed 7x: Salience 1.00 (back to maximum!)

Becomes "hot memory": access_count ≥10 + last_accessed <7 days

Why it matters: Frequently accessed info naturally stays strong without manual intervention!

Manual Reinforcement

4 Reinforcement Profiles:

Quick Refresh (+0.05)

When: Light reminder, brief review

Reinforce memory with Quick Refresh

Example: “Customer prefers email over phone”

Maintenance (+0.10)

When: Keep reference accessible

Reinforce memory with Maintenance

Example: Weekly reinforcement of API docs

Deep Learning (+0.15)

When: Actively studying/learning

Reinforce memory with Deep Learning

Example: Learning new technology weekly

Emergency (+0.25)

When: Critical info must persist

Reinforce memory with Emergency

Example: Production incident procedures

Reinforcement Strategies

Critical Info

Goal: Never let it fade

Strategy:

1. Store as Reflective (slowest decay)
2. Emergency profile (+0.25) immediately
3. Monthly reinforcement schedule

Result: Effectively permanent

Examples: Security protocols, core docs, strategic decisions

---

Access Patterns: Hot, Warm, Cold

System automatically categorizes memories based on access:

🔥 Hot

Definition: access_count ≥10 AND last_accessed <7 days

Effect:

• Surfaces first in searches
• High ranking boost
• Auto-reinforcement

Examples:

• Current feature docs
• Active customer context
• Sprint code patterns

🌡️ Warm

Definition: access_count ≥3 AND last_accessed <30 days

Effect:

• Good search ranking
• Available when needed
• Decaying slowly

Examples:

• Last month’s retrospective
• Recent customer issues
• Occasional code patterns

❄️ Cold

Definition: All other memories

Effect:

• Lower search ranking
• Natural decay continues
• Pruning candidate if salience less than 0.1

Examples:

• Old meeting notes
• Resolved bugs (months ago)
• Outdated temporary notes

Admin Panel: Stats widget shows hot/warm/cold breakdown. Hot memories = what your team values most!

---

Pruning: Automatic Cleanup

Trigger: Salience less than 0.1 after 90 days inactivity

Process:

1. Nightly job scans candidates
2. Checks salience threshold (default: 0.1)
3. Checks last accessed (>90 days)
4. Soft deletes (recoverable 30 days)
5. Hard deletes after grace period

Why automatic: Keeps knowledge base focused and performant without manual maintenance!

Configuration

Default (recommended):

• Threshold: 0.1 salience
• Inactivity: 90 days
• Grace period: 30 days

Can customize:

• Lower threshold (prune more)
• Higher threshold (keep more)
• Disable auto-pruning (manual only)

Admin Panel:

• View candidates before deletion
• Manual prune with custom thresholds
• Recover soft-deleted within 30 days

Manual Pruning

Via AI assistant:

Prune memories with salience below 0.1

Via API:

curl -X POST https://api.ulpi.io/api/v1/memories/prune \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{"threshold": 0.1, "dry_run": true}'

Dry run first! Preview before actually pruning.

---

Real-World Strategies

Set It & Forget It

Approach: Minimal intervention

1. Store naturally as you work
2. Let sectors handle decay
3. No manual reinforcement
4. Trust access patterns

Best for: Personal knowledge, small teams

Result: 80% benefit, 20% effort

---

FAQ

Can I disable decay entirely?

No, but you can make memories effectively permanent:

1. Use Reflective (693-day half-life)
2. Emergency reinforcement (+0.25) regularly
3. Result: ~5-10 year lifespan

Why not disable? Decay keeps knowledge relevant!

Forgot to reinforce something important?

Don’t worry! Automatic reinforcement from access keeps it alive.

Admin panel shows memories approaching pruning threshold.

How often to manually reinforce?

• Weekly: Critical docs, active learning
• Bi-weekly: Reference materials, conventions
• Monthly: Strategic decisions, long-term docs
• Never: Most memories! Let automatic handle it.

See decay curves?

Yes! Admin panel shows salience over time for each memory. Visualize decay curve and reinforcement events.

---

Next Steps

Cognitive Sectors

5 memory types and decay rates

Best Practices

Effective memory management

Workflows

Real-world decay & reinforcement

Admin Panel

Monitor memory health visually

---

Memories that matter stay strong. Noise fades away. Trust the system.