A local-first memory engine for coding agents. It captures what you and the
agent actually do, distills it into versioned long-term memory, and recalls it
automatically — fully offline, with , and
exposed to any tool through . Installs as a namespaced add-on and never
replaces your existing directory.
Local-first memory engine for coding agents (Claude Code/Cursor/Codex): SQLite+FTS5 retrieval, versioned bi-temporal memories, full-session capture, zero dependencies, MCP server
스타
83
도구
—
가동 시간
—
zero third-party dependencies
MCP
.claude/
English
What is CC Agent Brain
CC Agent Brain turns a stateless coding agent into one that remembers you across
sessions and across tools — all locally, no cloud, no API key.
user prompts / assistant replies / tool actions
↓ automatic hooks
append-only event stream (SQLite + FTS5) + readable Wiki (Markdown)
↓ distill
versioned long-term memory (layered, bi-temporal, auditable)
↓ BM25 + RRF retrieval, project DB first / global DB fallback
automatic recall in future sessions
↓ MCP server
the same memory in Claude Code, Cursor, Codex, Gemini CLI, Claude Desktop…
Why it exists (2026 memory landscape)
Agent memory is becoming standard infrastructure, but existing choices force a
trade-off: cloud memory APIs send your data away; heavy self-hosted stacks need
vector/graph databases; plain Markdown memory does not actually search or
version itself. CC Agent Brain occupies the middle:
Zero-dependency engine — Python stdlib sqlite3 + FTS5 (BM25), including
CJK bigram tokenization, so Chinese and English retrieval both work offline.
Real capture loop — user prompts and assistant replies / tool calls are
indexed (Stop hook reads the transcript; idempotent).
Versioned, trustworthy memory — facts are superseded/expired, never
silently overwritten; a timeline shows what changed and when.
MCP-native, tool-agnostic — one memory store for every MCP client.
Human-readable projection — Markdown Wiki and generated indexes keep you
in control; the DB is a single portable file.
Private by construction — common secret shapes are redacted at capture
time; personal data is git-ignored.
SQL audit: duplicates, broken chains, stale entries
6 Skills — onboarding, wiki update/search, memory distill/audit, material distill.
3 Agents — memory curator, wiki compiler, material distiller.
Memory model
Four layers — working (current task), episodic (timestamped events),
semantic (distilled facts/preferences), procedural (reusable workflows).
Each memory has status (active/superseded/expired/archived), validity window,
confidence, source and supersede chain. See docs/memory-design.md.
Optional E2E-encrypted private backup (local-first by default)
planned
Privacy
Open-source framework, no personal data inside. Everything runs on your machine;
there is no telemetry and no network call. memory/, wiki/, private/ and
*.db are git-ignored. See docs/product-analysis.md
for the product rationale.