Skip to main content

Kindling

Local-first memory and continuity for AI-assisted development.

Kindling gives AI coding tools — and you — a memory of what happened: tool calls, file edits, commands, errors, decisions, summaries, and pinned findings. It stores that context locally in SQLite, organises it into meaningful sessions, and retrieves it later with deterministic, explainable results.

You can use Kindling from the command line, as a daemon-backed Rust SDK, or as an embedded in-process library.

The model​

Kindling has a small, deliberate data model:

Events → Observations → Capsules → Retrieval
  • Observations are atomic, immutable records of something that happened: a tool call, a shell command, a file diff, an error, or a message.
  • Capsules are bounded groups of observations — a session, a workflow run — each with an intent and an open/close lifecycle.
  • Retrieval returns context in three predictable tiers: pins, the current summary, and ranked search hits, each with provenance.

Why Kindling​

Development sessions generate valuable context that is normally lost — chat history disappears, terminal scrollback clears, and reasoning fades. Kindling captures that context as it happens and makes it retrievable later, so you (or your assistant) can resume work without re-explaining everything.

It is built around three principles:

  • Local-first. Project memory lives on your machine in SQLite. No server, no account, no network round-trip.
  • Deterministic retrieval. Pins, the current summary, and ranked provider hits are returned in a predictable, explainable order — you get back what you stored, not what an opaque model guessed.
  • Tool-agnostic. Capture context manually with the CLI, automatically through an adapter (Claude Code, OpenCode, PocketFlow), or programmatically from your own Rust or TypeScript code.

A quick taste​

No project setup required — load sample memory and search immediately:

curl -fsSL https://raw.githubusercontent.com/eddacraft/kindling/main/install.sh | sh
kindling demo
kindling search "JWT"
kindling browse

Or initialise a project and capture manually:

kindling init
kindling log "JWT tokens expire after 15 minutes, not 1 hour"
kindling log --kind error "segfault in auth middleware after the upgrade"
kindling search "JWT"
Search Results for: "JWT"
==================================================

Candidates (1):

1. 5f1c… (score: 0.83)
   Type: observation
   Content: JWT tokens expire after 15 minutes, not 1 hour
   Time: 2026-06-20 10:31:04

How it fits together​

        You / AI coding tools
   ┌───────────┬───────────┬───────────────┐
   │   CLI     │  Adapters │   SDK crates  │
   │ kindling  │  (hooks)  │ kindling-*    │
   └─────┬─────┴─────┬─────┴───────┬───────┘
         └───────────┼─────────────┘
                     ▼
            kindling daemon / service
                     ▼
        SQLite (per project, WAL + FTS5)

The same store backs every entry point. A CLI invocation, a Claude Code hook, and a Rust client all read and write the same per-project database.

What Kindling is not​

  • Not a note-taking app — it captures structured observations, not freeform prose.
  • Not a chat logger — it stores knowledge and events, not transcripts.
  • Not a semantic black box — retrieval is mechanical full-text search with explainable ranking, not embeddings guessing at intent.

Where to go next​