# Cognitive Memory

Markdown-based AI memory system — MCP server, CLI, decay scoring, episodic logging. stdlib-only (no pip dependencies).

## Commands
- Run CLI: `cd /mnt/NV2/Development/cognitive-memory && uv run python client.py <command>`
- Wrapper: `claude-memory <command>` (installed at `~/.local/bin/claude-memory`)
- MCP server: `python3 mcp_server.py` (stdio JSON-RPC, registered in `~/.claude.json`)

## Architecture
- `common.py` — constants, YAML parsing, decay math, helpers (everything imports from here)
- `client.py` — `CognitiveMemoryClient` core API + CLI entrypoint (`if __name__`)
- `cli.py` — argparse CLI, called from `client.py`'s `__main__` block
- `edges.py`, `embeddings.py`, `analysis.py` — mixins composed into `CognitiveMemoryClient`
- `mcp_server.py` — standalone MCP server wrapping `CognitiveMemoryClient`
- Data lives at `~/.local/share/cognitive-memory/` (never in this repo)

## Style
- Flat module imports: `from common import ...`, `from client import ...` (no package prefix)
- Mixin pattern: `EdgesMixin`, `EmbeddingsMixin`, `AnalysisMixin` composed via multiple inheritance
- YAML frontmatter in memory files parsed manually (no PyYAML dependency)
- All paths via `pathlib.Path`

## IMPORTANT
- Zero external dependencies — stdlib only. Ollama/OpenAI are optional HTTP calls.
- `mcp_server.py` does `sys.path.insert(0, ...)` to find sibling modules — don't restructure without updating this
- `feature.json` has stale `files` entries from pre-extraction — update when modifying file list
- Skill layer at `~/.claude/skills/cognitive-memory/` (SKILL.md + SCHEMA.md) is separate — don't duplicate app code there