🧠

Smart Recall with Keyword Boost

First message loads all project memories with progressive disclosure (250-char summaries). "Remember" queries get boosted thresholds (0.75-0.85) and more results.

🤖

Self-Managing Agents

Agents are bootstrapped from cortex knowledge, evaluated on usage, updated with new insights, and retired when obsolete. Full lifecycle.

📚

Auto-Skill Discovery

Detects 50+ frameworks (FastAPI, Next.js, Flask, Go, Rust, etc.) from project markers and auto-generates slash-command skills with best practices. Manual /cortex discover adds web research.

🧬

Agent Context Injection

When any subagent spawns, domain-relevant memories are injected via SubagentStart hook. Agents start with knowledge, not from scratch.

🧹

Memory Hygiene

Daily automated dedup, path validation, recall tracking, and consolidation. Designed for long-running projects — never deletes based on age.

🔗

MCP Resources

Reference memories inline with @memory://project/name. 4 resource endpoints for on-demand context loading.

🛡️

Safety Guardrails

Audit trail for all operations, soft-delete with backup, content size limits, agent hard caps, path traversal protection.

🏷️

Auto Project Tagging

PreToolUse hook detects project from working directory and auto-tags every memory_store call. No manual project scoping needed.

📋

Global Behavioral Rules

Ships with config/CLAUDE.md and config/cortex-memory.md — global rules that teach Claude to always search memory before saying "I don't know".

💾

Persistent Agent Memory

All agents have memory: user or memory: project scope. They accumulate knowledge across sessions in their own MEMORY.md files.

Full Session Lifecycle

Session Start
  ├── cleanup.sh          prune stale data
  ├── agent_bootstrap.sh  create agents from cortex (daily)
  ├── memory_hygiene.sh   dedup + consolidate (daily)
  └── skill_discover.sh   detect tech stack + generate skills (weekly)

Every Message
  └── recall.sh           smart recall + keyword boost + progressive disclosure

Agent Spawned
  ├── agent_context_inject.sh  inject domain memories
  └── agent_track.sh           log to usage ledger

Context Compressed
  ├── compact_save.sh      extract memories + fleet management
  └── post_compact_save.sh extract from compressed summary

Every cortex Tool Call
  └── cortex_pretool_enrich.sh  auto-tag project + audit

Session Paused
  ├── learn.sh             prompt to save learnings
  └── fleet_eval_stop.sh   fleet health check

Session Ended
  └── session_end_cleanup.sh  save summary + cleanup

Hook Coverage

Hook Event	Script	What It Does
UserPromptSubmit	recall.sh	Smart recall with keyword boost + progressive disclosure
PreToolUse	cortex_pretool_enrich.sh	Auto-tag project + audit log for cortex ops
PostToolUse(Agent)	agent_track.sh	Track agent spawns in usage ledger
SubagentStart	agent_context_inject.sh	Inject domain memories into spawned agents
PreCompact	compact_save.sh	Extract memories + create/evaluate agents
PostCompact	post_compact_save.sh	Extract from compressed summary
SessionStart (x4)	cleanup + bootstrap + hygiene + skill_discover	Prune, create agents, dedup memories, generate skills
SessionEnd	session_end_cleanup.sh	Save session summary + cleanup
Stop (x2)	learn + fleet_eval_stop	Save learnings + fleet health check

Quick Install

1. Clone & Install

git clone https://github.com/digin1/cortex.git ~/.claude/skills/cortex
pip install chromadb

2. Initialize Database

python3 -c "import chromadb; c=chromadb.PersistentClient(path='$HOME/.claude/cortex-db'); c.get_or_create_collection('claude_memories'); print('OK')"

3. Configure Hooks

Copy the hooks configuration from the README into your ~/.claude/settings.json.

4. Install Global Behavioral Rules

cp ~/.claude/skills/cortex/config/CLAUDE.md ~/.claude/CLAUDE.md
mkdir -p ~/.claude/rules
cp ~/.claude/skills/cortex/config/cortex-memory.md ~/.claude/rules/cortex-memory.md

5. Restart Claude Code

🧠 85 memories (39 project, 17 feedback, 18 reference, 1 user) across 4 projects
🤖 10 agents (5 project + 5 global) | 5 spawns today | health 4.2/5
📚 3 skills (2 project + 1 global)

Safety Guardrails

Audit Trail

Every store, update, and delete logged to .cortex_audit.jsonl with timestamps

Soft Delete

Deleted memory content archived to audit log before removal

Content Limits

Max 5000 chars per memory, 200 total memories, agent hard caps

No Age Deletion

Designed for long-running projects. Memories never removed for being old

Path Protection

os.path.realpath() validation + directory whitelisting on all agent ops

Semantic Dedup

0.55 for agents, 0.15 for memories, 0.35 for hygiene merges