succ
Open Source|FSL 1.1
Cloud coming soon
succ

Semantic Understanding for Code Contexts

Persistent semantic memory for Claude Code. 14 consolidated MCP tools, knowledge graph, code search with tree-sitter AST, and working memory that never forgets.

View on GitHubGet Started
Open Source
|MCP Native|TypeScript

Core Capabilities

Long-term memory for your AI assistant

succ turns ephemeral AI conversations into persistent, interconnected knowledge. Every decision, pattern, and lesson is preserved, scored, and retrievable.

Working Memory
Pinned memories, priority scoring with confidence decay, and diversity filtering. Your AI assistant loads the most important context at every session start.
Hybrid Search
BM25 keyword search combined with semantic vector similarity. Search memories, brain vault docs, and source code with a single query.
Code Intelligence
Tree-sitter AST parsing for 13 languages. Extract functions, classes, and symbols. Search code by meaning, regex, or symbol type.
Knowledge Graph
Memories are linked with typed relationships. Community detection, centrality scoring, and automatic relationship enrichment build a web of understanding.
Brain Vault
Persistent knowledge base organized as Obsidian-compatible markdown. Indexed, searchable, and automatically enriched with wikilinks and metadata.
Session Continuity
Auto-loads context at session start, precomputes briefings during idle time, and hands off state between sessions. No memory loss between conversations.

MCP Tools

14 consolidated tools your AI assistant can use

Every tool is available via Model Context Protocol. Claude Code discovers and uses them automatically — no manual configuration needed.

Search

succ_search

Hybrid brain search (BM25 + semantic)

succ_search_code

Code search with regex + symbol filters

Indexing

succ_index

action="doc" — index docs into the brain vault

succ_index

action="code" — index source files with AST-aware chunks

succ_index

action="analyze" — generate docs from a file via LLM

succ_index

action="refresh" — reindex stale files and clean deletions

succ_index

action="symbols" — extract AST symbols (13 languages)

Memory

succ_remember

Store memories with auto-tagging and quality scoring

succ_recall

Hybrid search across memories with temporal filters

succ_forget

Delete memories by id, age, or tag

succ_dead_end

Record failed approaches to prevent repeats

Knowledge Graph

succ_link

Graph actions: create/delete/show/graph/auto/enrich/proximity/communities/centrality/export/cleanup/explore

PRD & Debug

succ_prd

action="generate|list|status|run|export" — full PRD pipeline

succ_debug

action="create|hypothesis|result|resolve|status|list" — structured debugging

Status & Config

succ_status

action="overview|stats|score" — status, token savings, AI readiness

succ_config

action="show|set|checkpoint_create|checkpoint_list" — config + backups

Web & Fetch

succ_web

action="quick|search|deep|history" — Perplexity-powered search

succ_fetch

Fetch and clean any URL to Markdown

Quick Start

Three steps to persistent memory

Get succ running in under a minute. No accounts required. API keys optional for advanced features.

01

Install and initialize

One command installs succ globally. Then run succ init in your project to set up MCP server, session hooks, and local storage.

terminal
# Install globally
npm install -g @vinaes/succ

# Initialize in your project
cd my-project
succ init

# That's it. succ is now:
#   MCP server registered in Claude Code
#   Session hooks installed
#   Daemon running in background
#   Ready to remember
02

Store knowledge naturally

Tell Claude to remember decisions, patterns, or lessons. succ scores quality, detects invariants, builds embeddings, and links related knowledge automatically.

claude code session
> "Remember: our API always uses cursor-based
   pagination, never offset-based"

succ_remember:
  id:          12847
  type:        decision
  tags:        [api, pagination, cursor]
  quality:     0.82
  invariant:   true  (detected "always/never")
  links:       2 related memories found
03

Recall with full context

succ searches memories, brain vault, and source code using hybrid BM25 + semantic search. Your AI assistant recalls decisions, patterns, and code with full relationship context.

claude code session
> "How did we handle pagination?"

succ_recall (3 results, 12ms):
  [decision] API uses cursor-based pagination
    score: 0.91 | pinned | tags: api, cursor
    linked: rate-limiting, auth-tokens

  [pattern] Cursor encoding uses base64
    score: 0.74 | tags: api, encoding

  [learning] Offset pagination breaks on
    concurrent writes — switched to cursor
    score: 0.68 | tags: api, migration

Architecture

Designed for developer trust

succ is built on principles that matter to developers: privacy, performance, and complete control over your data.

succ Architecture
Claude Code / AI Assistant
|
MCP Protocol (14 consolidated tools)
|
succ Daemon
Embedder
AST Parser
Graph
Search
Working Memory
Quality Scoring
Session Hooks
|
SQLite/PostgreSQL/Qdrant
Pluggable storage backends

Local-First, Scale-Ready

Starts with SQLite for zero-config setup. Scale to PostgreSQL + Qdrant when your project demands it.

Local Embeddings

ONNX Runtime with GPU auto-detection. Semantic search runs on your machine with no API calls required.

Knowledge Graph

Memories are linked in a knowledge graph with community detection, centrality scoring, and relationship enrichment.

Privacy by Default

Zero telemetry. All core features work offline. Optional web search and LLM features require explicit opt-in.

Background Daemon

Persistent daemon with REST API handles indexing, idle-time context precompute, and session handoff automatically.

Project Isolation

Each project gets its own memory space, brain vault, and knowledge graph. Global memories shared across all projects.

Updates

What changed recently

Fresh, dated release notes so search engines and AI agents know succ is alive and improving.

Feb 20, 2026

14 consolidated MCP tools

Merged 31 tools into 14 actions-first endpoints, added per-action profile gating and cleaner error surfaces for assistants.

Feb 3, 2026

Web search budgets + history

Daily spend guardrails, citation logging, and searchable web-history audit for Perplexity-powered calls.

Jan 18, 2026

Working-memory refresh

Better session handoff, pinned memories, and priority decay to keep Claude Code context sharp across long projects.

FAQ

Answers for humans & crawlers

Short, structured answers help search engines and LLMs surface the right facts about succ.

Do I need an API key?

Core memory, indexing, and graph features run fully local. Web search and cloud embeddings are opt-in via succ_config(action="set").

How many MCP tools are there now?

14 consolidated tools, each with actions. See the Tools section for action names you can call from Claude Code.

Is my code uploaded anywhere?

No telemetry. Everything stays on your machine unless you enable web search or a remote embedding provider.

Which IDEs work?

Claude Code with MCP is first-class. Any MCP-capable agent can call succ tools via the manifest and llms.txt hints.

For AI Agents

Everything an agent needs, visible

We expose crawlable entry points so Claude, ChatGPT, and other MCP-capable agents can self-serve: sitemap, llms.txt, and clear tool descriptions with action names.

View toolsllms.txt

Discovery

llms.txt and ai.txt at /.well-known point crawlers to the sitemap and allowed paths.

MCP surface

14 consolidated tools with action parameters; assistants learn them from on-page descriptions and metadata.

Config hints

Use succ_config(action="set") to set API keys or budgets; web budgets prevent runaway spend for autonomous agents.

Community

Built in the open, shaped by developers

succ is source-available under the FSL 1.1 (Functional Source License). We welcome contributions of all kinds, from bug reports to new features to documentation improvements.

Documentation

Help improve guides, tutorials, and API docs.

Discussions

Share ideas, report issues, and request features.

Code Contributions

Fix bugs, add features, and improve the codebase.

Contribute on GitHubJoin Discussion

Released under the Functional Source License 1.1 (FSL 1.1).

Support the project

If succ has been useful in your workflow, you can support its continued development. Every contribution, no matter the size, helps keep this project alive.

Click any address to copy. Thank you for your generosity.

Your AI forgets everything.
Fix that.

One command. 14 consolidated MCP tools. Persistent memory that survives every session, every context window, every restart.

Get StartedView Source

Works with Claude Code out of the box. No configuration needed.