cc-intel

                _       _       _
   ___ ___     (_)_ __ | |_ ___| |
  / __/ __|____| | '_ \| __/ _ \ |
 | (_| (_|_____| | | | | ||  __/ |
  \___\___|    |_|_| |_|\__\___|_|

Archived, August 2026. Claude Code ships its own memory now, writing facts to MEMORY.md as a session runs. I've stopped maintaining this. The code stays up for anyone who wants to read it.

Your Claude Code sessions are losing memory. cc-intel fixes that.

You spend an hour making architectural decisions with Claude Code. Then the context window fills up, compaction kicks in, and those decisions vanish. Next session? Claude starts from scratch, contradicts itself, and you repeat the same conversations.

cc-intel reads your session transcripts, uses Claude to extract what matters, and writes it to MEMORY.md so nothing gets lost.

---

What it does

Command	What it does
projects	Pick a project -- auto-finds the latest session to analyze
snapshot	Extract decisions, constraints, artifacts, and tasks
preserve	Same as snapshot, but writes results into MEMORY.md
risk	Check how full your context window is before compaction hits
status	Show MEMORY.md health and section budget usage

All commands are run as cc-intel <command>. By default they auto-discover your latest session from ~/.claude/projects/, or you can pass a specific session file with --input.

---

Install

This is a CLI tool -- install it globally so the cc-intel command is available everywhere:

npm install -g cc-intel

API key setup

cc-intel uses Claude Haiku to extract knowledge from your sessions. You need an Anthropic API key.

1. Get one at console.anthropic.com/settings/keys
2. Add it to your shell:

echo 'export ANTHROPIC_API_KEY=sk-ant-api03-your-key-here' >> ~/.zshrc
source ~/.zshrc

(On bash, use ~/.bashrc instead. On Windows, see the instructions shown when you run any command without a key set.)

Cost is roughly $0.001 per extraction.

Your key stays on your machine. cc-intel reads it from the environment variable at runtime and sends it only to the Anthropic API. It is never stored to disk, logged, or transmitted anywhere else. This is the same BYOK (bring your own key) pattern used by every major CLI tool.

---

Usage

Extract a snapshot

cc-intel snapshot

Reads the latest session from your current project and prints a structured breakdown:

  # Session Snapshot
  --------------------------------------------------

  ## Project Goal (2)
    - Build a REST API with Express and TypeScript
    - Add JWT authentication for all protected routes

  ## Key Decisions (5)
    - Use PostgreSQL with Drizzle ORM instead of Prisma
    - Store sessions in Redis for horizontal scaling
    ...

  ## Constraints (3)
    - API must return responses under 200ms at p99
    ...

  ## Implementation Artifacts (4)
    - Created src/routes/auth.ts with login/register endpoints
    ...

  ## Open Tasks (2)
    - Add rate limiting middleware
    - Write integration tests for auth flow

  --------------------------------------------------
  Extracted 16 entries from 85 messages

Preserve to MEMORY.md

cc-intel preserve            # write to MEMORY.md
cc-intel preserve --dry-run  # preview without writing

Extracts knowledge and merges it into your project's MEMORY.md with deduplication. Existing entries are preserved; duplicates are skipped. Budget limits keep the file from growing unbounded.

Check compaction risk

cc-intel risk

Estimates how full your context window is and classifies the risk:

# Context Risk Assessment

Risk Level: **HIGH**

Utilization: 72.3% [##############......]
Estimated tokens: 144,600 / 200,000
Messages: 127

Exits with code 1 when risk is High or Critical -- useful in CI or git hooks.

Check MEMORY.md health

cc-intel status

Shows budget utilization per section:

  MEMORY.md Status
  --------------------------------------------------
  Path: /Users/you/.claude/projects/.../memory/MEMORY.md
  Total: 142/200 lines (71%)

    Pinned Essentials: 45/80 [########.......] 56%
    Index Links: 12/40 [####...........] 30%
    Recent Decisions: 52/60 [#############..] 87%

Browse projects

cc-intel projects        # interactive picker
cc-intel projects --json # machine-readable list

Shows all Claude Code projects, lets you pick one, then runs snapshot/risk/preserve on it.

---

How it works

1. Parse -- reads native Claude Code session files (JSONL) from ~/.claude/projects/
2. Pre-filter -- scores messages by signal patterns (decisions, constraints, artifacts, TODOs) and picks the top ~60 messages
3. Extract -- sends the filtered messages to Claude Haiku, which identifies the actual knowledge worth preserving
4. Deduplicate -- compares new entries against existing MEMORY.md using Dice coefficient similarity
5. Write -- merges into MEMORY.md with atomic file writes and advisory locking

The pre-filtering step uses heuristic signal detection to reduce what gets sent to the LLM. This keeps cost under $0.001 per extraction while preserving extraction quality.

---

Session auto-discovery

When you run any command without specifying a file, cc-intel:

1. Finds the git root of your current directory
2. Looks in ~/.claude/projects/<project>/ for session files
3. Uses the most recently modified one

You can also pass a file explicitly:

cc-intel snapshot path/to/session.jsonl
cc-intel snapshot --json -o snapshot.json

Supported formats: native Claude Code JSONL (auto-detected), simple JSONL ({"role":"human","content":"..."} per line), and markdown (Human: / Assistant: prefixes).

---

MEMORY.md budget

MEMORY.md has strict line limits to stay useful inside Claude's system prompt:

Section	Default limit
Pinned Essentials	80 lines
Index Links	40 lines
Recent Decisions	60 lines
Total	200 lines

When a section overflows, excess content moves to a topic file and a link replaces it in MEMORY.md.

---

Configuration

Create .cc-intelrc.json in your project root or home directory:

{
  "maxContext": 150000,
  "logLevel": "debug",
  "riskThresholds": {
    "medium": 0.5,
    "high": 0.75,
    "critical": 0.9
  },
  "memoryBudget": {
    "maxLines": 200,
    "sectionLimits": {
      "pinnedEssentials": 80,
      "indexLinks": 40,
      "recentDecisions": 60
    }
  }
}

Environment variable overrides:

• CC_INTEL_LOG_LEVEL -- debug, info, warn, error
• CC_INTEL_MAX_CONTEXT -- max context token count

---

Library API

cc-intel exports its core functions for programmatic use:

import {
  parseSession,
  discoverLatestSession,
  segmentSession,
  extractSnapshot,
  parseMemoryDocument,
  mergeIntoMemory,
  assessRisk,
  loadConfig,
} from 'cc-intel';

See docs/api.md for the full API reference.

---

Development

git clone https://github.com/tornike14/cc-intel.git
cd cc-intel
npm install
npm run build
npm run test
npm run lint
npm run typecheck

---

Privacy and security

• BYOK model -- you provide your own Anthropic API key. cc-intel never stores, logs, or transmits it anywhere except the Anthropic API
• Local processing -- session parsing, scoring, deduplication, and memory management all run locally on your machine
• No telemetry -- cc-intel sends nothing home. No analytics, no tracking, no phone calls
• Open source -- every line of code is auditable. MIT licensed

License

MIT