Executive Summary

A Self-Evolving Memory System That Actually Learns

This system transforms passive data storage into an active learning engine. Inspired by Andrej Karpathy's LLM Knowledge Base architecture, it combines pattern extraction, Hebbian link detection, ACT-R cognitive modeling, and Ebbinghaus forgetting curves to create a memory system that improves over time.

System Architecture

Tier 1: Redis (Hot Cache)

127.0.0.1:6379

kilo:search:* — Search results (1hr TTL)
kilo:entry:* — Memory entries (24hr)
kilo:stats:* — Aggregations (5min)
shared:infra:* — Cross-agent data (12hr)

Tier 2: MariaDB (Primary)

127.0.0.1:3306

kilo — 1,209 memory entries
aoede — 15,507+ agent entries
claude — Knowledge library
infra_secrets — Credentials
8 databases total

Tier 3: Obsidian (Knowledge)

127.0.0.1:7654

wiki/ — Main documentation
kilo/ — Kilo-specific knowledge
harry/ — Harry's knowledge
raw/ — Research materials
338+ files across 4 vaults

Tier 4: Karpathy Learning

Self-improving engine

Pattern Extractor — 1,200+ entries analyzed
Hebbian Links — 396+ tag connections
ACT-R Activation — 18 types ranked
Feedback Loop — Outcome tracking
Qdrant — Vector search (L2 layer)

The 5 Learning Components

1

Pattern Extractor

Analyzes all memory entries to identify patterns in successful vs. failed actions. Categorizes by event_type and tags to understand what works.

# Example: What tags lead to success?
patterns["successful"]["by_tag"]["redis"] = 23
patterns["successful"]["by_tag"]["deployment"] = 15
patterns["successful"]["by_tag"]["obsidian"] = 31
          

2

Hebbian Link Detection

"Cells that fire together wire together." Detects co-occurring tags and creates weighted links between them — the system learns that redis and cache often go together.

# Example: Hebbian links (top 5)
{"from": "redis", "to": "cache", "weight": 47}
{"from": "memory", "to": "karpathy", "weight": 38}
{"from": "deployment", "to": "one-page-html", "weight": 29}
{"from": "obsidian", "to": "documentation", "weight": 26}
{"from": "infra", "to": "127.0.0.1", "weight": 21}
          

3

ACT-R Activation

Based on John Anderson's ACT-R cognitive theory. Successful patterns "float to top" through activation calculation: A = A_base + ln(1/time) + success_bonus.

# Example: ACT-R activation ranking
[{"type": "action", "activation": 4.2},
 {"type": "feature", "activation": 3.8},
 {"type": "learning", "activation": 3.1},
 {"type": "solution", "activation": 2.9},
 {"type": "fix", "activation": 2.4}]
          

4

Ebbinghaus Forgetting

Old failures fade away over time. Entries older than 14 days with event_type 'issue' or 'failed' are automatically deprioritized — the system "forgets" what didn't work.

Threshold: 14 days
Affected types: issue, error, failed, fix-failed

5

Feedback Loop

Tracks outcomes of actions. When you record "success" or "fail", it updates the system to suggest better actions in the future.

from karpathy_learn import FeedbackLoop
fb = FeedbackLoop()
fb.record_outcome("redis cache setup", "success")
fb.record_outcome("obsidian api", "fail")

# Get suggestions based on past success
recs = get_recommendations("cache redis")
          

Data Flow

Query

→

Redis

hot cache

→

MariaDB

primary

→

Obsidian

knowledge

→

Qdrant

vector search

→

Karpathy

learn

Write

←

Redis

←

MariaDB

←

Memory.py

Memory Layers (L1–L4)

Layer	Technology	Latency	Purpose
L1	Redis	0ms	Identity, rules, configs
L2	Qdrant	5-50ms	Semantic vector search
L3	MariaDB	5-10ms	Structured entries (1,200+ records)
L4	Obsidian	50-200ms	Rich content vault (338+ files)

Multi-Agent Memory

Kilo Agent

• 4+ Redis keys (search, entry, stats, obsidian)
• 1,209 entries in kilo database
• 191 sessions tracked
• Full R/W to MariaDB
• Access to wiki/kilo/ vaults

Harry Agent

• 23+ Redis keys (infra, services, creds)
• Separate entries in separate DB
• Access to wiki/harry/ vault
• shared:infra:* accessible to all

Shared Memory

Both agents can read shared:infra:* keys — when Kilo discovers 127.0.0.1 has MariaDB+Redis, Harry doesn't need to re-discover. The knowledge is shared.

Action Breakdown

1,209

Total Actions

191

Sessions

1,200+

Entries Analyzed

8

MariaDB Databases

Top Action Types

action

209

feature

182

obsidian

173

correction

138

learning

104

session-end

98

system

67

discovery

47

Usage

# Log to memory

python3 memory.py log action "Deployed feature" --tags feature,deploy

# Search memory

python3 memory.py search redis

# Access Obsidian

python3 obsidian_access.py search "workflow"

# Run learning cycle

python3 karpathy_learn.py

# Get recommendations

from karpathy_learn import get_recommendations

recs = get_recommendations("redis setup")