Kindex

The memory layer Claude Code doesn't have.

Kindex does one thing. It knows what you know.

It's a persistent knowledge graph for AI-assisted workflows. It indexes your conversations, projects, and intellectual work so that Claude Code never starts a session blind. Available as a free Claude Code plugin (MCP server) or standalone CLI.

Memory plugins capture what happened. Kindex captures what it means and how it connects. Most memory tools are session archives with search. Kindex is a weighted knowledge graph that grows intelligence over time — understanding relationships, surfacing constraints, and managing exactly how much context to inject based on your available token budget.

Install as Claude Code Plugin

Two commands. Zero configuration.

pip install kindex[mcp]
claude mcp add --scope user --transport stdio kindex -- kin-mcp
kin init

Claude Code now has 30 native tools: search, add, context, show, ask, learn, link, list_nodes, status, suggest, graph_stats, graph_merge, dream, changelog, ingest, tag_start, tag_update, tag_resume, remind_create, remind_list, remind_snooze, remind_done, remind_check, remind_exec, mode_activate, mode_list, mode_show, mode_create, mode_export, mode_import, mode_seed.

Or add .mcp.json to any repo for project-scope access:

{ "mcpServers": { "kindex": { "command": "kin-mcp" } } }

Install as CLI

pip install kindex
kin init

With LLM-powered extraction:

pip install kindex[llm]

With reminders (natural language time parsing):

pip install kindex[reminders]

With everything (LLM + vectors + MCP + reminders):

pip install kindex[all]

Why Kindex

Context-aware by design

Five context tiers auto-select based on available tokens. When other plugins dump everything into context, Kindex gives you 200 tokens of executive summary or 4000 tokens of deep context — whatever fits. Your plugin doesn't eat the context window.

Knowledge graph, not log file

Nodes have types, weights, domains, and audiences. Edges carry provenance and decay over time. The graph understands what matters — not just what was said.

Operational guardrails

Constraints block deploys. Directives encode preferences. Watches flag attention items. Checkpoints run pre-flight. No other memory plugin has this.

Cache-optimized LLM retrieval

Three-tier prompt architecture with Anthropic prompt caching. Stable knowledge (codebook) is cached at 10% cost. Query-relevant context is predicted via graph expansion and cached per-topic. Only the question pays full price. Transparent — kin ask just works better and cheaper.

Team and org ready

.kin inheritance chains let a service repo inherit from a platform context, which inherits from an org voice. Private/team/org/public scoping with PII stripping on export. Enterprise-ready from day one.

In Practice

A 162-file fantasy novel vault — characters, locations, magic systems, plot outlines — ingested in one pass. Cross-referenced by content mentions. Searched in milliseconds.

$ kin status
Nodes:     192
Edges:     11,802
Orphans:   3

$ time kin search "the Baker"
# Kindex: 10 results for "the Baker"

## [document] The Baker - Hessa's Profile and Message Broker System (w=0.70)
  → Thieves Guild, Five Marks, Thieves Guild Operations

## [person] Mia and The Baker (Hessa) -- Relationship (w=0.70)
  → Sebastian and Mia, Mia -- Motivations and Goals

0.142 total

$ kin graph stats
Nodes:      192
Edges:      11,802
Density:    0.3218
Components: 5
Avg degree: 122.94

192 nodes. 11,802 edges. 5 context tiers. Hybrid FTS5 + graph traversal in 142ms.

Getting Claude to Actually Use It

Installing the MCP plugin gives Claude the tools. But Claude won't use them proactively unless you tell it to. Kindex ships with a recommended CLAUDE.md block that turns passive tools into active habits:

# Print the recommended directives
kin setup-claude-md

# Or auto-append to your global CLAUDE.md
kin setup-claude-md --install

This adds session lifecycle rules (start/during/segment/end), explicit capture triggers (discoveries, decisions, key files, notable outputs), and search-before-add discipline. The difference between "Claude has a knowledge graph" and "Claude actively maintains a knowledge graph" is this block.

The SessionStart hook (kin setup-hooks) reinforces these directives at the start of every session with a "Session directives" block that reminds Claude to use kindex MCP tools throughout the session.

What gets captured

With the directives active, Claude will:

• Search the graph before starting work and before adding nodes
• Add discoveries, decisions, key files, notable outputs, and new terms as they emerge
• Link related concepts when connections are found
• Learn from long files and outputs via bulk extraction
• Tag sessions to track work context across conversations
• Remind with actions for deferred tasks (shell commands or headless Claude invocations)

Actionable Reminders

Reminders can carry shell commands and/or natural-language instructions. When due, the daemon executes them automatically — simple commands run directly, complex tasks launch headless claude -p. A Stop hook guard blocks Claude from exiting when actionable reminders are pending.

# Kill a cloud instance in 1 hour (but download results first)
kin remind create "Kill vast.ai instance" --at "in 1 hour" \
  --action "vastai destroy instance 12345" \
  --instructions "Download results from /workspace/ before killing"

# Manual trigger
kin remind exec --reminder-id <id>

Dream — Knowledge Consolidation

Kindex dreams. After each Claude Code session, a detached background process runs fuzzy deduplication, auto-applies pending suggestions, and strengthens edges between nodes that share domains. Like memory consolidation during sleep — replay, strengthen important paths, prune noise.

# See what would happen (no changes)
kin dream --dry-run

# Run full consolidation
kin dream

# Fast path: dedup + suggestions only
kin dream --lightweight

# Include LLM-powered cluster summarisation
kin dream --deep

# Fork and return immediately (used by Stop hook)
kin dream --detach --lightweight

Three triggers: manual CLI, periodic cron (step 11 of kin cron), and automatic detached subprocess on Claude Code session exit. File locking prevents concurrent cycles. The detached process uses start_new_session=True to survive Claude Code's exit.

Conversation Modes

Modes are reusable conversation-priming artifacts that induce a processing mode in an AI session. Based on research showing that induced understanding outperforms direct instruction by 5.4x, and that 15 tokens of mode-setting capture 98.8% of achievable priming benefit.

Five built-in modes: collaborate, code, create, research, chat. Create custom modes from any session and export them for team sharing (PII-free).

# Seed default modes
kin mode seed

# Activate a mode — outputs the priming artifact
kin mode activate collaborate

# Create a custom mode
kin mode create debug-session \
  --primer "We're hunting a bug. Precision over speed..." \
  --boundary "Show your reasoning chain. Name assumptions." \
  --permissions "Speculate about root causes freely."

# Export for team sharing (PII-stripped)
kin mode export collaborate > collaborate.json

# Import a teammate's mode
kin mode import their-mode.json

Modes are not instructions — they're state inductions. A primer establishes how to think, a boundary defines what quality means, and permissions state what's allowed. The AI shifts processing mode rather than following a checklist.

Quick Start

# Add knowledge (with optional tags)
kin add "Stigmergy is coordination through environmental traces" --tags biology,coordination

# Search with hybrid FTS5 + graph traversal
kin search stigmergy
kin search coordination --tags biology   # filter results by tag

# Ask questions (with automatic classification)
kin ask "How does weight decay work?"

# Get context for AI injection
kin context --topic stigmergy --level full

# List and filter by tags
kin list --tags python,ml              # nodes tagged with both
kin list --type concept --tags ai      # combine type and tag filters

# Track operational rules
kin add "Never break the API contract" --type constraint --trigger pre-deploy --action block

# Check status before deploy
kin status --trigger pre-deploy

# Ingest from all sources
kin ingest all

# Session tags — named work context handles
kin tag start auth-refactor --focus "OAuth2 flow" --remaining "tokens,tests"
kin tag segment --focus "Token storage" --summary "Flow design done"
kin tag resume auth-refactor   # context block for new session
kin tag end --summary "All done"

# Reminders — never forget, never nag
kin remind create "standup" --at "every weekday at 9am" --priority high
kin remind create "reply to Kevin" --at "in 30 minutes" --priority urgent
kin remind list
kin remind snooze --reminder-id <id> --duration 1h
kin remind done --reminder-id <id>

.kin/ Directory & Inheritance

Projects use .kin/ directories that encode their communication style, engineering standards, and values. Teams inherit from orgs. Repos inherit from teams. The knowledge graph carries the voice forward.

~/.kindex/voices/acme.kin             # Org voice (downloadable, public)
    ^
    |  inherits
~/Code/platform/.kin/config           # Platform team context
    ^
    |  inherits
~/Code/payments-service/.kin/config   # Service-specific context

# payments-service/.kin/config
name: payments-service
audience: team
domains: [payments, python]
inherits:
  - ../platform/.kin/config

The .kin/ directory is the standard location for all kindex project artifacts:

• .kin/config — project metadata (voice, domains, audience, inheritance)
• .kin/index.json — graph snapshot for git tracking

The payments service gets Acme's voice principles, the platform's engineering standards, AND its own domain context. Local values override ancestors. Lists merge with dedup. Parent directories auto-walk when no explicit inherits is set.

Old-style .kin files (plain YAML) are auto-upgraded to .kin/config on first access.

See examples/kin-voices/ for ready-to-use voice templates.

Architecture

SQLite + FTS5          <- primary store and full-text search
  nodes: id, title, content, type, weight, audience, domains, extra
  edges: from_id, to_id, type, weight, provenance
  fts5:  content synced via triggers

Retrieval pipeline:
  FTS5 BM25 --+
  Graph BFS --+-- RRF merge -- tier formatter -- context block
  (vectors) --+                   |
      |                   full | abridged | summarized | executive | index
      |
  Embedding providers (configurable):
      local (sentence-transformers) | openai | gemini

LLM cache tiers (kin ask):
  Tier 1: codebook (stable node index)     <- cached @ 10% cost
  Tier 2: query-relevant context           <- cached per-topic @ 10% cost
  Tier 3: user question                    <- full price, tiny

Reminders:
  reminders table (SQLite)    <- separate from knowledge graph
  Time parsing:  dateparser (NL) + dateutil.rrule (recurrence) + cronsim (cron)
  Channels:      system (macOS) | slack | email | claude (hook) | terminal
  Daemon:        launchd/cron adaptive interval -> check due -> notify -> auto-snooze
  Scheduling:    adaptive tiers (>7d=daily, >1d=hourly, >1h=10min, <1h=5min, none=disabled)
  Actions:       shell commands run directly | complex tasks launch claude -p
  Stop guard:    blocks session exit when actionable reminders pending

Dream (kin dream):
  Modes:         lightweight (<5s) | full (non-LLM) | deep (claude -p clusters)
  Triggers:      CLI | cron step 11 | Stop hook (detached, start_new_session=True)
  Dedup:         difflib.SequenceMatcher, 4-char title bucketing, 0.95 merge / 0.85 suggest
  Consolidation: suggestion auto-apply, domain edge strengthening, cluster summarisation
  Safety:        fcntl.flock exclusion, protected types skip, provenance tracking

Three integration paths:
  MCP plugin --> Claude calls tools natively (search, add, learn, remind, ...)
  CLI hooks  --> SessionStart / PreCompact / Stop lifecycle events
  Adapters   --> Entry-point discovery for custom ingestion sources
  Code       --> ctags + cscope + tree-sitter structural analysis

Node Types

Knowledge: concept, document, session, person, project, decision, question, artifact, skill

Code Intelligence

Ingest repository structure with kin ingest code --directory .:

• Module nodes (artifact) — one per source file with structural summary: classes, public functions, signatures, imports
• Symbol nodes (concept) — one per class/interface/type with method signatures
• Edges — imports (depends_on), inheritance (implements), containment (context_of), call graph (relates_to)
• Three extraction tiers — ctags (100+ languages), cscope (C/C++ cross-refs), tree-sitter (AST call graphs)
• Incremental — file hashing skips unchanged files on re-ingest

Code structure lives in the same graph as your decisions, watches, and constraints. Search finds both what calls a function and what broke last time someone changed it.

Operational: constraint (invariants), directive (soft rules), checkpoint (pre-flight), watch (attention flags)

CLI Reference (49 commands)

Core

Knowledge Management

Graph Analytics

Operational

Ingestion & External Sources

Infrastructure

Configuration

Config is layered like git — global defaults, then global config, then local config. Each layer deep-merges over the previous, so you only set what you want to override.

Use kin config set --global llm.enabled true for global settings, or kin config set llm.model claude-sonnet-4-6 for project-local.

data_dir: ~/.kindex

llm:
  enabled: false
  model: claude-haiku-4-5-20251001
  api_key_env: ANTHROPIC_API_KEY
  cache_control: true              # Prompt caching (90% savings on repeated prefixes)
  codebook_min_weight: 0.5         # Min node weight for codebook inclusion
  tier2_max_tokens: 4000           # Token budget for query-relevant context

embedding:
  provider: local                  # local, openai, or gemini
  # model: ""                      # empty = provider default
  # api_key_env: ""                # empty = provider default (OPENAI_API_KEY / GEMINI_API_KEY)
  # dimensions: 0                  # 0 = provider default (384 / 1536 / 3072)

budget:
  daily: 0.50
  weekly: 2.00
  monthly: 5.00

project_dirs:
  - ~/Code
  - ~/Personal

defaults:
  hops: 2
  min_weight: 0.1
  mode: bfs

reminders:
  enabled: true
  check_interval: 300            # 5 min base interval
  adaptive_scheduling: true      # adjust interval based on nearest reminder
  min_interval: 300              # floor for adaptive scheduling
  default_channels: [system]     # system, slack, email, claude, terminal
  snooze_duration: 900           # 15 min default snooze
  auto_snooze_timeout: 300       # auto-snooze after 5 min inaction
  idle_suppress_after: 600       # suppress if idle > 10 min
  channels:
    slack:
      enabled: false
      webhook_url: ""
    email:
      enabled: false
      smtp_host: ""
      to_addr: ""

Development

make dev          # install with dev + LLM dependencies
make test         # run 980 tests
make check        # lint + test combined
make clean        # remove build artifacts

License

MIT