Kinthic — the local-first agent that remembers. Across sessions, on your machine.

---

Kinthic is a local-first AI agent built on the Silex memory engine — not a chat log with vectors bolted on. It persists facts in SQLite, indexes them semantically, and maps relationships in an epistemic graph so it can recall what it learned and how things connect.

AGPL-3.0 · data lives in ~/.kinthic/ · loopback gateway by default

Quickstart

curl -fsSL https://kinthic.openyf.dev/install.sh | bash
kinthic init          # wizard: provider, skills, optional Telegram/MCP
kinthic               # interactive agent session

Windows: use WSL2. Step-by-step: docs/quickstart.md · docs.kinthic.com

Try memory in 30 seconds (isolated demo):

python benchmarks/memory_recall/demo.py

Architecture

flowchart TB
  subgraph inputs [Inputs]
    CLI[CLI / TUI]
    TG[Telegram / Discord]
    MCP[MCP clients]
  end
  Loop[Cognitive Loop]
  AMAC[A-MAC admission]
  SQLite[(SQLite + FTS)]
  Vectors[(Chroma vectors)]
  Graph[(Epistemic graph)]
  Recall[Hybrid RRF recall]
  LLM[LLM context]

  CLI --> Loop
  TG --> Loop
  MCP --> Loop
  Loop --> AMAC
  AMAC --> SQLite
  AMAC --> Vectors
  Loop --> Graph
  SQLite --> Recall
  Vectors --> Recall
  Graph --> Recall
  Recall --> LLM

Loading

At the core:

• A-MAC — admission control filters low-quality memories before they persist
• Dual-store — SQLite commits first; vectors follow with reconciliation on startup
• Hybrid recall — blends recency, importance, keyword search, and semantic similarity (RRF)
• Epistemic graph — beliefs with confidence that decay and update over time

CLI essentials

Command	Purpose
kinthic init	First-run wizard (provider, skills, channels, MCP)
kinthic	Start interactive agent session
kinthic daemon start	Background supervisor (channels, proactive goals)
kinthic web	Local dashboard (metrics, graph, skills)
kinthic mcp serve --stdio	Expose Silex memory to Claude Desktop / Cursor
kinthic mcp print-config	Paste-ready MCP client JSON
kinthic skills install <name>	Install workflow skills from KinthicHub
kinthic data migrate --from openclaw	Import legacy agent state
kinthic data backup / restore	Export or restore ~/.kinthic
kinthic benchmark recall	Run memory recall benchmark

Full reference: docs/cli_reference.md

Features

• Persistent memory — hybrid retrieval, consolidation, tamper-evident rows, self-healing vector sync
• Skills — Markdown workflows in ~/.kinthic/skills/; install from catalog or URL
• MCP — use Kinthic as the memory layer for external agents (silex_recall, silex_remember, …)
• Channels — Telegram and Discord adapters
• Trajectory export — SFT / GRPO datasets from your agent's own runs
• Local-first — your brain is a folder on disk; nothing phones home by default

Memory recall benchmark

Reproducible needle-in-haystack test: 12 facts, 500 distractor memories, facts backdated 21 days (seed 42):

Baseline (aged 21d)	Hit@5	Hit@12
No memory	0%	0%
Keyword only	42%	50%
Vector only	92%	92%
Kinthic hybrid	67%	92%

Hybrid matches vector at Hit@12 and beats keyword-only on paraphrased queries. Reproduce:

kinthic benchmark recall --seed 42

Methodology: docs/benchmarks/memory-recall.md · raw results: benchmarks/memory_recall/results/REPORT.md

Comparison

Category difference, not a feature checklist:

Product	Role	Kinthic's angle
OpenClaw	Channel gateway	Routes messages; Kinthic remembers across sessions
Hermes	Task agent	Runs workflows; Kinthic builds a persistent brain
Kinthic	Memory + cognition	Dual-store recall, graph beliefs, skills, MCP memory server

Migrate: kinthic data migrate --from openclaw|hermes — docs/migration.md

Documentation

• docs/quickstart.md — golden path install → run
• docs/cli_reference.md — all commands
• docs/mcp.md — MCP setup and Silex memory tools
• docs/benchmarks/memory-recall.md — benchmark methodology
• docs.kinthic.com — hosted docs

Security

Paranoid by design for an agent with access to your machine:

• Loopback gateway — API bound to 127.0.0.1 with local API key
• Tool approvals — high-risk tools (writes, shell) require explicit approval
• Sandboxed execution — terminal commands run in Docker when enabled
• Memory guard — injection filtering on memory writes

Read SECURITY.md before deployment.

Contributing & roadmap

• CONTRIBUTING.md
• docs/planning/ROADMAP.md
• Commercial licensing: contact via GitHub org openyfai