CodyMaster Brain: The Cognitive Architecture
Standard AI systems treat memory as an append-only log, inevitably leading to overloaded context, "goldfish memory", and recurring hallucinations. CodyMaster Brain addresses this by implementing a biomimetic memory architecture inspired by human cognitive processes. It doesn't just store information; it synthesizes, filters, and selectively forgets.
The Five-Tier Memory System
CodyMaster implements a unified memory pipeline with five distinct tiers, each serving a specific cognitive function — from millisecond reflexes to permanent structural understanding.
Tier 1: Sensory Memory
Transient, single-command context — forgotten immediately.
Temporary information within one command execution: "file X has 200 lines." Discarded instantly after use because it has no long-term value. Like hearing a phone number once — gone in seconds.
Tier 2: Working Memory (cm-continuity)
Active session state — persists 7 days.
Just as humans can only hold about 7 items in short-term memory, CodyMaster maintains strict working memory limits. Managed by the cm-continuity skill, this layer tracks the active task, recent context, mistakes made, and immediate goals via CONTINUITY.md.
- When to use: Automatically loaded at session start, saved at session end.
Tier 3: Long-Term Memory (Context Backbone v5)
Experience-based learnings — StorageBackend abstraction (SQLite/OpenViking).
When bugs are fixed or patterns discovered, key insights are pushed to the Context Backbone v5. Depending on your .cm/config.yaml, this abstracts over two engines:
- SqliteBackend: Local WAL-mode SQLite database using FTS5 for lightning-fast localized text searches.
- VikingBackend: A robust REST-based semantic vector memory powered by the OpenViking daemon (
http://localhost:1933).
Both engines support Token Budget Enforcement—pre-allocating a precise max tokens budget limit for learnings per prompt. Only learnings relevant to the current module are loaded per task via progressive indexing (L0/L1/L2) and semantic filtering.
Tier 4: Semantic Memory (OpenViking)
Deep document retrieval — permanent, true vector search.
When context exceeds working memory capacity, CodyMaster interacts with the OpenViking daemon /search and /abstract endpoints (or qmd as a local fallback). This yields true vector similarity across massive codebases and documentation—even when query terms don't match exactly (e.g. "async timeout" matches "network latency spike"). Like a photographic reading memory — it finds patterns hidden in deep codebase semantics.
Tier 5: Structural Memory (cm-codeintell)
Real-time code architecture understanding — AST/CodeGraph.
cm-codeintell provides a 4-layer intelligence system that pre-indexes code structure for instant understanding:
| Layer | Name | Speed | Best For |
|---|---|---|---|
| 0 | Skeleton Index | <4s | Any project — zero dependencies |
| 1 | Code Graph | ~30s | 50+ files — call chains, impact analysis |
| 2 | Architecture | ~10s | 20+ files — visual module map |
| 3 | Smart Context | ~5s | Task-focused context packets |
Achieves 95% token compression by reading AST structure, not raw code. Like an innate sense of spatial awareness for your codebase.
The Cloud Brain (cm-notebooklm)
Beyond the 5-tier local pipeline, CodyMaster implements a Cloud Brain powered by Google's NotebookLM for cross-machine persistence:
- Master Brain (
codymaster): A global knowledge base storing cross-project meta-learnings and architecture patterns. - Project Brains (
project_id): Isolated NotebookLM instances for each project containing domain-specific PRDs, APIs, and business logic. - Wisdom Gatekeeper: Uses the Rule of 3 (recurring ≥ 3 times) to filter noise. Only qualified "Graduated Wisdom" is pushed to the global Master Brain.
- Audio Onboarding: Converts complex project knowledge into Spotify-style podcasts for new developers.
- Flashcard Learning: Auto-generated flashcards from the codebase help engineers quickly learn the stack.
The cm-notebooklm skill orchestrates this cloud sync via brain-sync.sh.
Biomimetic Mechanisms
CodyMaster Brain implements dynamic mechanisms to manage the lifecycle of information.
Smart Decay (The Forgetting Curve)
Standard AI systems hold onto outdated knowledge forever, leading to bugs when dependencies change. CodyMaster Brain implements a decay mechanism inspired by the Ebbinghaus Forgetting Curve. Old refactored code patterns and obsolete learnings literally "decay" and expire, keeping your active context clean and highly relevant.
Self-Healing Feedback Loops
When a bug leads to an infinite loop or a hallucination, CodyMaster Brain recognizes the pattern and "heals" its memory. The new Self-Healing AI Pipeline (cm-skill-health, cm-skill-evolution) continuously monitors success rates and auto-patches failing skills using a version DAG.
Progressive Loading & Budget Constraints
Instead of loading a massive data file on every request, the Context Backbone heavily relies on cm:// URIs and L0/L1/L2 progressive loading. For example, an initial hit fetches an L0 index (saving up to 96% of tokens), and only escalates to a full file read if deeper understanding is necessary. Combined with Token Budget Enforcement, this prevents context window overflow.
Integrating Core Memory Skills
To fully utilize the CodyMaster Brain, these core skills should be wired into your standard workflow:
| Skill | Role in Memory Architecture | Usage |
|---|---|---|
🧠 cm-continuity | Working Memory Engine | Run at the start and end of complex sessions to maintain state across boundaries without bloating the prompt. |
🔍 cm-deep-search | Semantic Retrieval | When the context exceeds the working memory capacity, this skill intelligently fetches relevant knowledge using BM25 or vector search. |
🏗️ cm-codeintell | Structural Intelligence | Pre-indexes codebase architecture (AST, CodeGraph) so AI understands code structure with 95% token compression. |
☁️ cm-notebooklm | Cloud Brain Sync | Synchronizes high-value project knowledge to Google NotebookLM for cross-machine persistence and audio/flashcard generation. |
🗄️ cm-dockit | Knowledge Organizer | Analyzes codebases and generates structured, LLM-readable knowledge bases (AGENTS.md, SOPs) that the brain can rapidly query. |
👑 cm-skill-mastery | Meta-Cognition | Helps the AI "think about thinking," knowing which memory structures to access and how to sequence them effectively. |
By mirroring human memory systems — with 5 cognitive tiers, cloud synchronization, smart decay, and self-healing loops — CodyMaster Brain turns the AI from a stateless code generator into a contextual engineering partner that learns, remembers, and grows across sessions and machines.