Continuous Claude: Give Claude Code a Memory That Survives Compaction

Claude Code compacts your conversation when the context window fills, and the nuance you built up over the session goes with it. Continuous Claude (CC-v3) — a 4k-star MIT-licensed setup — is built around that problem. Its mantra is "compound, don't compact": extract learnings automatically, then start each session fresh with the state already loaded.

Why This Skill Matters

This is not a single skill folder; it is a full environment the README counts as 109 skills, 32 agents, and 30 hooks installed into Claude Code. Three systems do the work. The continuity system keeps state in files: a markdown ledger per session topic under thoughts/ledgers/, and YAML handoff documents that a PreCompact hook auto-saves before compaction hits. The memory system runs PostgreSQL with pgvector in Docker and holds four tables — sessions, file claims, archival memory, and handoffs; when a session's heartbeat goes stale past five minutes, a daemon spawns a headless Claude run that extracts learnings from the session's thinking blocks. And the TLDR code analysis tool summarizes code in five layers, from AST through call graph, control flow, data flow, and program slicing — the README puts the summary at roughly 1,200 tokens against 23,000 raw, a 95% saving.

Hooks tie it together: on every prompt, a skill-activation hook injects a check telling Claude which skills and agents are relevant, ranked from CRITICAL to OPTIONAL, so you describe what you want instead of memorizing commands.

Installation

Prerequisites: Python 3.11+, the uv package manager, Docker (for PostgreSQL), and the Claude Code CLI. Clone the repository and run the 12-step setup wizard — always from the opc/ directory, where the pyproject.toml lives:

git clone https://github.com/parcadei/Continuous-Claude-v3.git
cd Continuous-Claude-v3/opc
uv run python -m scripts.setup.wizard

One note on naming: the README's later Installation section shows the pre-rename name continuous-claude in its clone URL; GitHub redirects it here, but Continuous-Claude-v3 is the repository's current name. The wizard backs up an existing .claude/ config before touching it, and the same wizard with --uninstall reverses the install while preserving your command history, MCP config, and API keys.

Real Workflow: Run a Bug Fix That Survives a Context Clear

Start with the router when you are not sure where a task belongs:

/workflow

It routes between Research, Plan, Build, and Fix. For a concrete bug, the /fix workflow chains the sleuth diagnostic agent through a premortem risk check, the kraken implementation agent, a test pass, and a commit:

/fix bug "login fails silently"

While you work, hooks track edits and index handoffs. When the session ends — or when compaction would wipe it — the state lives on: say "done for today" to create a handoff document, then open the next session with "resume work" and the ledger and handoff load automatically. Learnings the daemon extracted from previous sessions surface as MEMORY MATCH indicators when a new message is relevant to them.

Tips

• The core loop works without any paid API keys: the README lists continuity, TLDR analysis, local git operations, and the TDD workflow as functional with no keys at all. Perplexity and NIA keys are optional add-ons.
• The wizard offers copy and symlink install modes; symlink is the contributor-friendly one, since edits in ~/.claude/ and in the repo stay the same files.
• For brownfield projects, run /onboard after install — it analyzes the codebase and writes the initial continuity ledger.
• An optional math system adds /math (SymPy, Z3, Pint) and /prove, a five-phase Lean4 formal-verification workflow the README credits with the first Lean formalization of the Sylvester–Gallai theorem.
• PostgreSQL can be remote: point CONTINUOUS_CLAUDE_DB_URL at any instance with the pgvector extension enabled.

When Not to Use This

This is a full-environment install, not a drop-in skill folder: it wants Docker plus PostgreSQL, and it rewrites your ~/.claude/ setup (archived first, but still a takeover). The README itself warns there are "likely too many agents" and that consolidation is a v4 goal. And the repository's latest push was in January 2026 as of this writing, so check its current state before building your daily workflow on it.

See the leaderboard for more skills.