v0.12.0 930+ tests passing MIT License

The memory layer Claude Code doesn't have.

A persistent knowledge graph for AI-assisted workflows. Your AI never starts a session blind.

Memory plugins capture what happened. Kindex captures what it means and how it connects.

Terminal
$ pip install kindex[mcp]
$ claude mcp add --scope user --transport stdio kindex -- kin-mcp
$ kin init
Kindex initialized. 37 MCP tools available.
Core Capabilities

Everything your AI assistant needs to remember

A weighted knowledge graph that grows intelligence over time — understanding relationships, surfacing constraints, and managing context injection.

Knowledge Graph

Not a log file. Nodes with types, weights, domains. Edges with provenance and decay. The graph understands what matters.

5 Context Tiers

From 100 to 4000 tokens. Auto-selects the right depth based on your available token budget. Never eats your context window.

Operational Guardrails

Constraints block deploys. Directives encode preferences. Watches flag attention items. Checkpoints run pre-flight.

Team & Org Ready

.kin inheritance chains from service to platform to org. Private/team/org/public scoping with PII stripping on export.

Blazingly Fast

Hybrid FTS5 + graph traversal. 192 nodes, 11,802 edges searched in 142ms. Built on SQLite for zero-config speed.

Git Integration

Auto-capture knowledge on commits. Git hooks install with one command. Your codebase history becomes searchable context.

Cache-Optimized LLM

Three-tier prompt architecture with Anthropic prompt caching. Stable knowledge cached at 10% cost. Only the question pays full price.

Session Tags

Named work context handles that replace resume files. Track focus, remaining items, and topic segments. Resume seamlessly in new sessions.

Task Management

Actionable work items linked to graph concepts. Tasks surface contextually via graph proximity. Priority, status, and completion tracking built in.

Reminders

Time-based triggers with shell commands or Claude instructions. Natural language scheduling. Snooze, execute, or complete. Fires at session start.

Watches

Ongoing attention flags for flaky tests, unstable APIs, tech debt. Set owner and expiry. Watches surface in every session's context automatically.

Conversation Modes

Research-backed priming artifacts that induce processing modes in AI sessions. Five built-in modes. Create custom modes from sessions and export for team sharing.

Security

Your knowledge graph is protected

signet-eval gates every tool call with deterministic policy enforcement. Define rules for what your AI can and can't do with your knowledge.

signet-eval evaluates in ~2µs. Zero runtime overhead.

Learn more at signet.tools
policy.yaml
rules:
- name: protect-knowledge-graph
action: confirm
conditions:
- type: tool_name
values: [Bash, Write]
message: "Modifying files requires confirmation"
- name: allow-kindex-tools
action: allow
conditions:
- type: mcp_server
values: [kindex]
Smart Context

Five tiers of context, zero wasted tokens

When other plugins dump everything into context, Kindex auto-selects the right depth. 200 tokens of executive summary or 4,000 tokens of deep context — whatever fits.

Your plugin doesn't eat the context window. It adapts to it.

full
Session start, deep work~4,000 tokens
abridged
Mid-session reference~1,500 tokens
summarized
Quick orientation~750 tokens
executive
Post-compaction re-injection~200 tokens
index
Existence check only~100 tokens
The Difference

Graph intelligence, not session archives

Other Memory Plugins

  • -Session archives with keyword search
  • -Flat text storage, no relationships
  • -Dumps everything into context window
  • -No concept of importance or decay
  • -Single-user, single-project only

Kindex

  • Weighted knowledge graph with typed nodes and edges
  • Relationships, provenance, and edge decay over time
  • 5 context tiers that adapt to your token budget
  • Weights, domains, and audiences prioritize what matters
  • Team/org inheritance with privacy scoping
Real World

Built for serious work

From novel writing to multi-service architectures, Kindex scales to your knowledge.

Creative Writing

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

Creative Writing
$ kin status
Nodes: 192
Edges: 11,802
Orphans: 3
$ kin search "the Baker"
# 10 results for "the Baker"
[document] The Baker - Hessa's Profile (w=0.70)
> Thieves Guild, Five Marks
0.142s total

Multi-Repo Engineering

Link knowledge across repositories with .kin inheritance chains. A service repo inherits from platform context, which inherits from org-wide standards. Every Claude session knows the full stack.

Multi-Repo Engineering
# Service inherits from platform and org
$ cat .kin
inherit:
- ../platform/.kin
- ~/.config/kindex/org.kin
$ kin context --budget 1500
[abridged] 1,487 tokens injected
3 constraints, 2 directives, 1 watch

Research & Analysis

Track research papers, API documentation, competitive analysis. Kindex maintains weighted connections between concepts so you can ask questions that span your entire knowledge base.

Research & Analysis
$ kin ask "What are the tradeoffs between RAG and fine-tuning?"
Based on 4 connected nodes:
RAG offers real-time knowledge with higher latency.
Fine-tuning bakes knowledge in but goes stale.
Related: [constraint] Always cite sources
Ecosystem

The shared memory of your entire agent ecosystem

Kindex isn't just for you. Every agent in the Exemplar stack queries the same knowledge graph.

Sentinel

Fixer agents query Kindex for prior incident patterns before attempting remediation

Pact

Code generation agents query for architectural decisions and constraints

Chronicler

Records operational events as queryable knowledge that enriches future sessions

Get Started

Three ways to install

Choose your preferred integration. Zero configuration required.

Two commands. Zero configuration. Claude Code gets 37 native tools instantly.

Claude Code Plugin
$ pip install kindex[mcp]
$ claude mcp add --scope user --transport stdio kindex -- kin-mcp
$ kin init
Kindex initialized. Knowledge graph ready.
37 MCP tools available:
search, add, context, show, ask, learn, ingest,
link, list_nodes, status, suggest, graph_stats,
graph_heal, graph_merge, changelog,
tag_start, tag_update, tag_resume,
task_add, task_list, task_done,
remind_create, remind_list, remind_snooze,
remind_done, remind_check, remind_exec,
watch_add, watch_list, watch_resolve,
mode_activate, mode_list, mode_show,
mode_create, mode_export, mode_import, mode_seed
Tools & Commands

37 MCP tools. 49 CLI commands.

Comprehensive control over your knowledge graph, from both AI-assisted and manual workflows.

search

Find relevant knowledge by query (filter by --tags)

add

Add knowledge with type, weight, and tags

context

Get tiered context for current session

show

Display a specific node's full details

ask

Ask questions across your knowledge graph

learn

Extract and store knowledge from text

link

Create typed edges between nodes

list_nodes

Browse nodes with filters (type, status, tags, audience)

status

Graph overview and health

suggest

AI-powered suggestions for what to capture

graph_stats

Density, components, degree distribution

graph_heal

Diagnose graph health: orphans, bridges, fading nodes

graph_merge

Merge duplicate nodes, moving edges and archiving source

changelog

What changed, by whom, when

ingest

Ingest from GitHub, git, files, sessions, code (ctags/cscope/tree-sitter)

tag_start

Start a named session tag for work context

tag_update

Update, segment, pause, or end a session tag

tag_resume

Resume a session with full context injection

task_add

Create actionable work items linked to concepts

task_list

List and filter tasks by status and priority

task_done

Mark tasks completed with optional summary

remind_create

Create time-based reminders with natural language parsing

remind_list

List all active reminders

remind_snooze

Defer a reminder to a later time

remind_done

Mark a reminder as completed

remind_check

Run the reminder check cycle for due items

remind_exec

Manually trigger a reminder's action

mode_activate

Activate a conversation mode -- returns priming artifact

mode_list

List available modes (built-in and custom)

mode_show

Show mode details: primer, boundary, permissions

mode_create

Create a custom mode from primer/boundary/permissions

mode_export

Export a mode as PII-free portable artifact

mode_import

Import a mode from a portable artifact

mode_seed

Seed the five default conversation modes

watch_add

Flag items needing ongoing attention (flaky tests, tech debt)

watch_list

List active watches with status

watch_resolve

Archive a watch when the issue is resolved

Under the Hood

Simple architecture, powerful results

SQLite-Backed Graph

Zero-config storage with FTS5 full-text search. Your knowledge lives in a single portable file at ~/.kindex.

Typed Nodes & Edges

Nodes carry type, weight (0.0-1.0), domain tags, and audience scope. Edges track provenance and naturally decay, keeping your graph fresh.

Context Engine

Hybrid BFS + FTS5 traversal assembles context from graph relationships. Five tiers auto-select based on available token budget for optimal context injection.

Three-Tier LLM Cache

Codebook (stable node index) and query-relevant context are cached via Anthropic prompt caching at 10% cost. Graph edges predict which nodes to pre-cache per topic.

Graph Stats
$ kin graph stats
Nodes: 192
Edges: 11,802
Density: 0.3218
Components: 5
Avg degree: 122.94
Node types:
document: 78 (40.6%)
person: 42 (21.9%)
concept: 31 (16.1%)
constraint: 18 ( 9.4%)
directive: 14 ( 7.3%)
other: 9 ( 4.7%)
Weight distribution:
0.9-1.0 ████████ 23
0.7-0.9 ████████████████ 54
0.5-0.7 ████████████████████ 68
0.3-0.5 ████████████ 34
0.0-0.3 ████ 13
Research

Built on peer-reviewed foundations

Part of something bigger

Ready to give your AI a memory?

Install in under a minute. Free, open source, MIT licensed. Your knowledge stays local.