HCC Memory — hierarchical project memory (trace, findings, wisdom) for Claude Code
npx claudepluginhub chosenx-gpu/hcc-pluginHierarchical Cognitive Caching — project-level memory system for Claude Code. Tracks work, distills experience, builds verified knowledge across sessions.
Hierarchical Cognitive Caching — project-level memory for Claude Code.
HCC Memory is a Claude Code plugin that gives your AI assistant persistent, structured memory across sessions. Instead of losing context when a session ends, HCC captures execution traces, distills them into validated findings, and promotes the best ones to cross-project wisdom. It follows a three-layer model (trace → findings → wisdom) with progressive validation levels.
This repo ships a plugin marketplace catalog at .claude-plugin/marketplace.json. Install like any third-party marketplace plugin:
In Claude Code, add the marketplace (GitHub owner/repo):
/plugin marketplace add CHOSENX-GPU/HCC-Plugin
Install the plugin (plugin id hcc-memory @ marketplace id hcc-plugin):
/plugin install hcc-memory@hcc-plugin
Apply changes:
/reload-plugins
Optional: choose install scope (user / project / local) from /plugin → Discover → select hcc-memory, instead of typing the install command.
From a terminal (same naming as in-app):
claude plugin install hcc-memory@hcc-plugin --scope user
See Discover and install plugins for details.
To hack on the plugin or run without the marketplace flow:
git clone https://github.com/CHOSENX-GPU/HCC-Plugin.git
cd HCC-Plugin
claude --plugin-dir "$(pwd)"
On Windows (PowerShell), pass the full path to HCC-Plugin instead of $(pwd).
Initialize memory in your project directory (not inside the plugin repo):
/hcc-memory:init
Start a task:
/hcc-memory:plan "Run VKI LS89 cascade simulation"
Work normally. Every 5 tool uses, the 5-Action Rule will remind you to log progress. When you learn something valuable, use /hcc-memory:promote.
| Command | Description |
|---|---|
/hcc-memory:init | Initialize memory system in current project |
/hcc-memory:plan | Start a new task with description |
/hcc-memory:log | Log a trace entry (usually automatic via 5-Action Rule) |
/hcc-memory:promote | Distill trace entries into structured findings |
/hcc-memory:complete | Complete current task, archive trace, review for upgrades |
/hcc-memory:recover | Recover context after a session break |
/hcc-memory:search | Search memory for relevant findings and wisdom |
/hcc-memory:status | Show current memory system status |
/hcc-memory:compact | Audit memory for redundancy and staleness |
/hcc-memory:export | Export domain-scoped findings for knowledge sharing |
Findings are validated progressively:
| Level | Meaning | Confidence |
|---|---|---|
| syntax | Program ran without crashing | Low |
| numerical | Solution converged (⚠ not physics) | Medium |
| physical | Compared against experimental/analytical data | High |
| methodology | Validated across multiple case types | Highest |
Important: A numerical validation means "it converged" — not "it's correct." Always verify physical plausibility before trusting a numerically-validated finding.
After every 5 tool uses, the plugin reminds the agent to log its recent work. This ensures continuous trace capture and prevents context loss.
project/
├── memory/
│ ├── trace.md # Rolling execution log
│ ├── findings/ # L2 structured learnings
│ │ ├── _index.md # Auto-generated index
│ │ └── F-OF-EF-a1b2c3.md # Individual finding
│ ├── wisdom/ # L3 verified knowledge
│ │ ├── _index.md
│ │ └── W-OF-EF-a1b2c3.md
│ ├── tasks/ # Task tracking
│ │ └── _active.md
│ ├── sessions/ # Archived trace sessions
│ └── _export/ # Sanitized exports
├── .hcc/
│ ├── config.yaml # Plugin configuration
│ └── state.json # Runtime state
└── .gitignore # Auto-updated