Rekindle

For Claude Code users who lose time re-explaining project context every session.

npx rekindle init

Your AI forgets everything between sessions. Rekindle fixes that.

Rekindle is an MCP continuity engine that solves session orientation, not just storage. Orient at session start, capture at session end, survive mid-session compaction. All local, all SQLite, zero API keys.

v0.5.1: fixes for unfilled user cards and init --global, building on v0.5's portable user card. Release notes.

Quick Start

Requires Node.js 22 or newer.

npx rekindle init

This creates .rekindle/ in your project with a SQLite database, identity template, continuity policy, project registry, captures directory, and transcript directory. Then add the MCP server config for your client:

One command registers the server:

claude mcp add rekindle -- npx -y rekindle

Or add it to ~/.claude.json by hand:

{
  "mcpServers": {
    "rekindle": {
      "command": "npx",
      "args": ["-y", "rekindle"]
    }
  }
}

Enable PreCompact protection (captures context before mid-session compaction):

npx rekindle setup-hooks

Enable session-start orientation delivery — the budgeted orientation packet arrives automatically at startup, resume, /clear, and /compact, so the model re-orients at every context boundary without being asked:

npx rekindle setup-delivery

Both hooks are opt-in; plain init never installs either. npx rekindle init --with-hooks --with-delivery does everything in one line.

Without --global, both commands write the project-local .claude/settings.local.json, so the hooks run only for sessions started in that directory. With --global, they write the user-scope ~/.claude/settings.json, which Claude Code applies in every project on the machine; register the server at the same scope so it is present wherever the hooks fire: claude mcp add --scope user rekindle -- npx -y rekindle (plain claude mcp add is local to the current project).

Session-start delivery resolves the hook's real working directory through .rekindle/projects.json. Exact roots match by default; nested directories inherit only when an entry sets "include_descendants": true. An unmapped or ambiguous workspace is reported as UNKNOWN PROJECT and never receives a checkpoint or context record from another project. REKINDLE_PROJECT remains an explicit override, and receipts record which resolution rule won.

.rekindle/policy.json controls transcript handling, the startup byte budget, and mandatory versus optional packet sections. Mandatory sections are atomic: if they do not fit, Rekindle emits a bounded failure manifest instead of a partial anchor. Optional sections are omitted whole and recorded in the receipt. Set "transcript_mode": "excluded" to prevent SessionStart from reading transcript files or paths and to mark transcript coverage N/A in the orientation score. Set "background_writes_allowed": false to make the automatic PreCompact hook exit without creating a capture; manual capture remains a separate, explicit action.

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "rekindle": {
      "command": "npx",
      "args": ["-y", "rekindle"]
    }
  }
}

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "rekindle": {
      "command": "npx",
      "args": ["-y", "rekindle"]
    }
  }
}

Then fill in .rekindle/identity.md and paste the boot instructions into your project's CLAUDE.md.

Session 1 stores. Session 2 remembers. Session 10 anticipates.

Related MCP server: claude-session-continuity-mcp

The Problem (43 Sessions of Data)

Over 43 sessions, we measured what an AI assistant failed to load at session start:

Existing memory tools (Mem0, Letta, Zep) optimize for retrieval accuracy: can the AI find what it stored? That's necessary but not sufficient. None of them address whether the AI loaded the right context for this session, or whether it can detect what it missed.

Rekindle solves session orientation: loading identity, recent context, memory health, and missing-context warnings before the assistant starts work.

See docs/gap-analysis.md for the full research dataset.

What It Does

Boot: orient at session start

boot_report runs an orientation pipeline before any work begins:

boot_report
  +-- Read identity document (who am I working with?)
  +-- Scan memory stats (what do I know?)
  +-- Find latest checkpoint (where did we leave off?)
  +-- Read last transcript (what actually happened?)
  +-- Surface open loops (what needs follow-up?)
  +-- Surface PreCompact captures (what survived compaction?)
  +-- Detect gaps (what am I missing?)
  +-- Calculate orientation score (how oriented am I?)
  --> "Carrying forward: [context loaded, gaps identified, score: 80/100]"

Survive the Long Middle: PreCompact capture (v0.3)

Mid-session compaction destroys reasoning chains, failed approaches, relational texture, and tone. The PreCompact hook fires automatically before compaction and saves what would otherwise be lost:

PreCompact hook fires
  +-- Parse JSONL transcript (last N messages)
  +-- Write raw Markdown capture (.rekindle/captures/)
  +-- Write structured JSON snapshot (decisions, open loops, files)
  +-- Update manifest for cheap listing
  --> boot_report surfaces captures on next session start
  --> end_session warns if captures exist but weren't reviewed

Three read modes control token cost:

• summary — one paragraph, cheap

• structured — decisions/loops/warnings, moderate

• raw — full transcript excerpt, expensive (only when needed)

Capture: close the loop at session end

end_session stores structured continuity records — not just a summary:

All records stored with type, source, and session_id metadata. Next boot_report loads the checkpoint automatically.

Between sessions: search and manage

Categories: preference lesson context relationship general

Carry you between agents: the user card (v0.5)

identity.md tells the agent who it is. The user card tells any agent who you are: how to talk to you, your boundaries, your projects, what to avoid. It's one plain Markdown file that you own and edit; agents suggest changes, you decide.

Start one from the template:

npx rekindle card template ~/user-card.md     # never overwrites an existing file

Or draft one from what your agent already knows, review it, and save what you accept as your card:

mkdir -p ~/card-draft
npx rekindle card compile --out ~/card-draft  # writes card-draft.md and card-worksheet.md
cp ~/card-draft/card-draft.md ~/user-card.md  # after you've reviewed and edited the draft

Then point this store at it:

npx rekindle card path ~/user-card.md

card path only records where the card is. An agent receives the card through Rekindle's orientation output: the SessionStart delivery hook (npx rekindle setup-delivery) in Claude Code, the Cursor adapter, or a boot_report call from any other MCP client.

• Opt-in. If no card is configured, the card feature adds nothing: no card section and no user_card receipt field.

• Whole or nothing. With the default policy, the card is the first optional section after the required ones. If it doesn't fit the startup budget it's left out, never cut off partway.

• Diagnostics stay out of the packet. What happened to the card (included, dropped, missing, unreadable, empty, unfilled) is recorded in the receipt's user_card field, which holds only the source and status, never the card text or path.

• compile drafts, you decide. It selects records whose type is preference or constraint, or whose category is preference (project-less by default, plus any projects you name, each project line dated), your identity file's "What I Know About My Person" section, and any reviewed files you pass with --from, which it copies verbatim. It reads at most 2,000 records, newest first. It never writes over your card. Its detection of conflicts, secrets and stale entries is incomplete, so the draft is a candidate list for you to review, not a cleaned card. Details.

• One card, many agents. Point every store you use at the same file. Agents that can't read your machine (cloud agents) need their own copy; see docs/user-card.md.

• Trust limit. Any agent with ordinary file access can edit the card file. Rekindle doesn't prevent that; keep the card somewhere you review.

REKINDLE_USER_CARD overrides the configured path; a blank value counts as unset.

Why not just CLAUDE.md?

A static file is passive. Your AI reads it, but it can't search it, rank it, track what's been retrieved, or tell you what's missing. Rekindle adds:

• Search — full-text with importance-weighted ranking

• Structure — category and project scoping across memories

• Orientation — proactive context loading at boot, not just on-demand retrieval

• Gap detection — flags missing identity, empty categories, stale data

• Scoring — transparent checklist so you know how oriented the AI is

• Session capture — structured close with checkpoints, decisions, and open loops

• Compaction survival — PreCompact captures preserve what summaries flatten

Release Highlights

v0.5.1

• An unfilled user card is left out (#38): a card holding only the shipped template and plain one-line comments is no longer delivered, and the receipt says unfilled.

• init --global leaves ~/.gitignore alone (#26): project init still adds .rekindle/ to that project's .gitignore.

v0.5.0

• User card (#32): one Markdown file about the user that any agent can read. rekindle card template, card path, card compile, and init --card. Opt-in: if no card is configured, the card feature adds no packet section and no receipt field (the other changes below still apply). See above

• Transcript discovery stays in the current workspace (#29): automatic discovery could return another project's transcript. It now matches Claude Code's folder naming and only accepts a transcript that records this workspace as its working directory

• capture_now keeps a transcript with its own session (#28): with only transcript_path, the session id comes from the file name and discovery is skipped; with only session_id, a discovered transcript is used only if it belongs to that session; with both, both are used as given. A blank value counts as omitted, and a resulting id that is empty or hidden is refused

• Captures are tagged with the resolved project (#30): captures use the same project resolution as session start. Behavior change: a capture made where no project resolves is now untagged (project: null, with a warning) instead of being labelled with the folder name

• An unfilled identity template no longer scores (#31): the shipped template earns no identity points and shows an identity_unfilled warning; the text is still delivered. Behavior change: the score line now reads "Identity document filled in", and fresh installs score 20 points lower until identity.md is filled in

v0.4.1

• --global hooks fire everywhere — setup-hooks --global, setup-delivery --global and init --global --with-* now write ~/.claude/settings.json (user scope) instead of ~/.claude/settings.local.json, which Claude Code reads only for sessions started in the home directory (#25). Project-local setup is unchanged. Each command prints the file it wrote

• Unresolved-workspace warning names the fix — when the session-start packet cannot map the workspace to a project it now says that no checkpoint was loaded and points at the two remedies: add the workspace to the registry it read (.rekindle/projects.json for a project store, ~/.rekindle/projects.json for a --global store), or set REKINDLE_PROJECT. The warning does not expose the absolute registry path

• Init output matches this README — init leads with claude mcp add rekindle -- npx -y rekindle (claude mcp add --scope user ... under --global, so the server is registered wherever the global hooks fire), lists both opt-in hooks (with --global when used), and prints ~/.rekindle/... paths under --global

v0.4.0

• Node.js 22 or newer required — Node 20 reached end of life in April 2026 and is no longer supported; CI runs on Node 22, 24 and 26

• Policy-aware session start — .rekindle/policy.json controls transcript handling (transcript_mode: captured, tail-read-only, or excluded), the startup byte budget, mandatory versus optional packet sections, and whether background hooks may write (background_writes_allowed)

• Project-safe delivery — .rekindle/projects.json maps workspaces to projects; an unmapped or ambiguous workspace is reported as UNKNOWN PROJECT and never receives another project's checkpoint; REKINDLE_PROJECT is an explicit override

• Transcript filename withheld in tail mode — in tail-read-only mode the packet shows the transcript tail but not the file's name

• Atomic session writes — a session row and its audit row commit together or not at all

• Clean production audit — npm audit --omit=dev found zero vulnerabilities (production dependencies)

• Smaller package — the npm tarball drops from about 1.9 MB to about 110 KB; evidence and design documents stay on GitHub

v0.3.3

• Version-consistent protocol metadata — the MCP initialize response derives its version from the shipped package metadata, preventing release-version drift

• Package-page accuracy — the README shipped to npm identifies the current release before the tag and package are created

• Expanded automated tests, plus a packed-artifact check that compares MCP metadata to the installed package version

v0.3.2

• One-command delivery install — npx rekindle setup-delivery (or init --with-delivery) configures the SessionStart hook opt-in: idempotent, preserves other tools' hooks, refuses corrupted settings files

• 147 automated tests

v0.3.1 — "Five Measured Gates"

• Session-start delivery — rekindle session-start emits a budgeted orientation packet via the SessionStart hook at startup, resume, /clear, and /compact

• Budgeted packets, truthful receipts — packets cap at 8,000 valid UTF-8 bytes with an in-packet truncation marker; receipts attest emission only and never claim model visibility

• Desktop-safe storage — storage root never derives from the spawn point (Claude Desktop spawns MCP servers at /); explicit resolution order, fail-loud

• Dual-channel guidance — workflow guidance rides both tool descriptions and MCP instructions, drift structurally impossible

• Cursor adapter — session-start --client cursor with whitelist stdin parsing; email and workspace paths never reach receipts

• Measured, not assumed — every claim above is backed by a published measurement (evidence, spike results)

v0.3.0 — "Survive the Long Middle" added the PreCompact capture system, open loops, and review tracking — v0.3.0 release notes

CLI Commands

Install from Source

git clone https://github.com/Skitchy/rekindle.git
cd rekindle
npm install
npm run build
node dist/init/cli.js init

The setup-hooks command writes this to .claude/settings.local.json (or ~/.claude/settings.json with --global):

{
  "hooks": {
    "PreCompact": [
      {
        "matcher": "auto",
        "hooks": [
          {
            "type": "command",
            "command": "npx rekindle precompact-capture",
            "timeout": 60
          }
        ]
      },
      {
        "matcher": "manual",
        "hooks": [
          {
            "type": "command",
            "command": "npx rekindle precompact-capture",
            "timeout": 60
          }
        ]
      }
    ]
  }
}

The hook receives session context on stdin (session_id, transcript_path, cwd, hook_event_name) and writes captures to .rekindle/captures/.

Storage root resolution. All Rekindle entry points (server, PreCompact hook) resolve the directory holding .rekindle/ through one rule, in order:

1. REKINDLE_BASE_DIR, if set — explicit always wins

2. Derived from REKINDLE_DB_PATH, when it points at a canonical <base>/.rekindle/db/ layout

3. An existing .rekindle/ in the current working directory (never when cwd is the filesystem root)

4. An existing .rekindle/ in your home directory

5. Otherwise: your home directory — never the spawn point

Rules 3 and 5 exist because some hosts (e.g. Claude Desktop) spawn MCP servers at cwd=/; a spawn point is not a storage location. If storage cannot be created, the server exits with a message naming the fix instead of a stack trace.

• All data is local. Nothing is sent to external servers.

• No network calls. The MCP server communicates via stdio. No HTTP, no telemetry, no analytics.

• Transcripts contain conversation text. Do not enable transcript capture if your sessions contain secrets or credentials.

• Hook installation is opt-in. Both the capture hook (setup-hooks) and the delivery hook (setup-delivery) must be requested explicitly, by command or by flag. Plain init never installs either.

• SQLite database is a regular file. Not encrypted. Use OS-level disk encryption if needed.

• Project-local .rekindle/ is gitignored. Project init adds it to that project's .gitignore; init --global leaves ~/.gitignore unchanged.

• boot_report reads local files. Paths are not sandboxed. Only use with MCP clients and prompts you trust.

Compatibility

"Full delivery" means the orientation packet arrives automatically at session boundaries and the model demonstrably sees it — measured with canary probes at both the receipt layer and the model layer, not assumed. Details and evidence: compatibility spike results.

Claude Code: session-start orientation (opt-in)

npx rekindle setup-delivery

writes this to .claude/settings.local.json (or ~/.claude/settings.json with --global):

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup|resume|clear|compact",
        "hooks": [
          { "type": "command", "command": "npx rekindle session-start", "timeout": 60 }
        ]
      }
    ]
  }
}

The packet is capped at 8,000 valid UTF-8 bytes — measured: when hook output exceeds the host's limit, the model sees only the leading portion, with no error surfaced. If sections are dropped to fit the budget, an in-packet marker says so, and the receipt in .rekindle/receipts/session-start.jsonl records exactly what was emitted without ever claiming the model saw it.

Cursor: session-start orientation (opt-in)

Cursor's hook system can deliver the budgeted orientation packet at session start, measured working in the v0.3.1 compatibility spike. Setup is manual and opt-in — Rekindle never installs hooks without being asked. Add to .cursor/hooks.json in your project:

{
  "version": 1,
  "hooks": {
    "sessionStart": [ { "command": "rekindle session-start --client cursor" } ]
  }
}

Privacy: Cursor's hook payload includes your account email and workspace paths. The adapter treats that payload as personal by default: it extracts only the session ID and workspace root (used in-process for storage resolution), and neither the raw payload, the email, nor any path is ever written to receipts or any other artifact. Background agents are bypassed by default (truthfully receipted); opt in with REKINDLE_ORIENT_BACKGROUND_AGENTS=1.

rekindle/
  src/
    index.ts          MCP server entry point
    server.ts         Server setup, tool registration (10 tools)
    storage/
      sqlite.ts       SQLite + FTS5, schema migration, sessions
    orientation/
      types.ts        OrientationResult, Gap, ScoreItem
      GapDetector.ts  Structural gap detection (8 codes)
      Scorer.ts       Orientation scoring (6 criteria, 100pts)
      OrientationService.ts   Orchestrator
      OrientationRenderer.ts  Markdown + JSON output
    captures/
      types.ts        CaptureEntry, StructuredSnapshot, HookInput
      CaptureManager.ts   Parse, capture, list, read, review tracking
      discover-transcript.ts  Auto-discover session transcripts
      precompact-capture.ts   CLI hook entry point
      capture-now.ts          Manual capture CLI
    tools/
      boot-report.ts  Orientation + open loops + capture awareness
      end-session.ts  Structured session close + capture warning
      list-captures.ts  List PreCompact captures
      read-capture.ts   Read captures in 3 modes
      capture-now.ts    Model-triggered manual capture
      store.ts search.ts list.ts delete.ts update.ts
    delivery/
      budget.ts       UTF-8 packet construction and atomic mandatory sections
      receipts.ts     Emission receipts (never claim model visibility)
      session-start.ts SessionStart hook adapter
      cursor.ts       Cursor hook adapter (privacy-whitelisted stdin)
      guidance.ts     Canonical workflow guidance, both channels
    init/
      cli.ts scaffold.ts setup-hooks.ts setup-delivery.ts templates/

Storage: SQLite + FTS5 via better-sqlite3. BM25 ranking boosted by importance. Typed records with type, source, session_id.

Transport: stdio (standard MCP). Works with Claude Code out of the box.

Tests

npm test

The suite runs in CI on Node 22, 24 and 26 and covers: storage CRUD + FTS5 ranking, orientation domain (gap detection, policy-aware scoring, project resolution, service, rendering), capture manager (parsing, limits, review tracking, formatting), delivery (atomic packet budget, receipts, guidance channels, Cursor privacy sentinels), the user card (path resolution, whole-or-nothing delivery, compile safeguards), hook setup for both hooks (schema, idempotency, corruption refusal), and MCP integration (all 10 tools plus package-derived server metadata).

Roadmap

Next: "It thinks in networks" — Spreading activation, semantic search via embeddings, gap analysis tooling, eval harness.

Contributing

See CONTRIBUTING.md, the changelog and the Code of Conduct.

License

MIT