claude-mem-sync

claude-mem-sync

Team memory sharing for claude-mem — sync AI memories across developers via git.

Official documentation

Table of Contents

• Why This Exists
• What It Does
• Step-by-Step Setup Guide
  • Step 1: Install Prerequisites
  • Step 2: Install the Claude Code Plugin
  • Step 3: Run the Setup Wizard
  • Step 4: Create the Shared Team Repository
  • Step 5: Set Up GitHub Actions (CI/CD)
  • Step 6: Your First Export
  • Step 7: Import Team Memories
  • Step 8: Enable Developer Profiles
  • Step 9: Enable Knowledge Distillation (optional)
  • Step 10: Launch the Dashboard
  • Step 11: Set Up Automatic Scheduling (optional)
• Configuration Reference
• CLI Reference
• Web Dashboard
• Developer Knowledge Profiles
• Knowledge Distillation
• Eviction & Scoring
• Destination Patterns
• Maintenance
• Security & Privacy
• Architecture
• Publishing a Release (maintainers)
• Troubleshooting
• License

Why This Exists

claude-mem gives Claude persistent memory across sessions, storing observations (decisions, bugfixes, discoveries) in a local SQLite database. But it’s designed for single-user, per-machine usage.

When a team works on the same project:

• Each developer has their own isolated memory database
• Developer A discovers a critical pattern — developers B through L never learn about it
• The same bugs get rediscovered, the same decisions re-debated
• There is no native team mode, no shared database, no automatic sync

claude-mem-sync bridges this gap with filtered, scored, deduplicated team memory sharing using git as the transport layer.

What It Does

Developer A                    GitHub (shared repo)              Developer B
┌──────────┐    export         ┌──────────────────┐    import   ┌──────────┐
│ claude-   │ ──────────────►  │ contributions/   │  ◄───────── │ claude-  │
│ mem.db    │    (filtered)    │   dev-A/         │  (merged)   │ mem.db   │
│           │                  │   dev-B/         │             │          │
│           │    import        │                  │   export    │          │
│           │ ◄──────────────  │ merged/          │ ──────────► │          │
│           │    (deduped)     │   latest.json    │  (filtered) │          │
└──────────┘                   └──────────────────┘             └──────────┘
                                    ▲
                               GitHub Action
                               (merge + dedup + cap)
                                    │
                          ┌─────────┴──────────┐
                          ▼                    ▼
                    ┌────────────┐      ┌────────────┐
                    │  profiles/ │      │ distilled/ │
                    │  per-dev   │      │ rules.md   │
                    │  metrics   │      │ kb.md      │
                    └────────────┘      └────────────┘
                    (deterministic)     (LLM-powered)

Features:

• Filtered export — only share what matters (by type, keyword, or tag)
• Intelligent eviction — scoring system prevents unbounded DB growth
• Deduplication — composite key dedup across developers
• Access tracking — PostToolUse hook tracks which memories Claude actually uses
• Cross-platform scheduling — automatic export/import via cron, launchd, or Task Scheduler
• PR review mode — optional human review before memories enter the shared repo
• Multi-provider support — GitHub, GitLab, and Bitbucket (including self-hosted)
• CI merge bot — templates for GitHub Actions, GitLab CI, and Bitbucket Pipelines
• Web dashboard — 9-tab dark-theme UI with charts, heatmaps, profiles, and analytics
• Rich analytics — type distribution, access patterns, developer contributions, observation scoring
• Developer knowledge profiles — per-dev metrics: knowledge spectrum, concept map, file coverage, temporal patterns, survival rate
• Knowledge distillation — LLM-powered extraction of CLAUDE.md rules and knowledge docs from team observations
• Team insights — knowledge gaps detection, concept coverage heatmaps, bus-factor risk analysis
• Configurable cleanup — automatic retention policy for old contribution files

Step-by-Step Setup Guide

This guide walks you through the complete setup of claude-mem-sync, from zero to a fully working team memory sharing pipeline. Follow each step in order.

Step 1: Install Prerequisites

Before you start, make sure you have these tools installed on your machine.

1.1 — Install Bun (recommended)

You need one of these runtimes. Bun is recommended because it’s faster and has built-in SQLite.

Option A: Install Bun (recommended)

# macOS / Linux
curl -fsSL https://bun.sh/install | bash

# Windows (PowerShell)
powershell -c "irm bun.sh/install.ps1 | iex"

# Verify it works
bun --version   # Should print 1.x.x or higher

1.2 — Install Git

# macOS (Homebrew)
brew install git

# Ubuntu/Debian
sudo apt install git

# Windows — download from https://git-scm.com/

# Verify it works
git --version

1.3 — Install GitHub CLI (required for PR mode and repo creation)

# macOS (Homebrew)
brew install gh

# Ubuntu/Debian
sudo apt install gh

# Windows (winget)
winget install GitHub.cli

# Log in to GitHub
gh auth login

# Verify it works
gh --version

1.4 — Install claude-mem

claude-mem-sync reads from claude-mem’s database. You need claude-mem installed and already generating observations.

Follow the claude-mem installation guide to set it up.

After installation, verify you have a database file at:

• Default location: ~/.claude-mem/claude-mem.db

# Check the file exists
ls -la ~/.claude-mem/claude-mem.db

Don’t have observations yet? Use Claude Code for a few sessions first. claude-mem automatically records decisions, bugfixes, and discoveries as you work.

Step 2: Install the Claude Code Plugin

This single step installs both the plugin (PostToolUse hook for access tracking) and the CLI tool (mem-sync command) automatically.

# Add the claude-mem-sync marketplace
claude plugin marketplace add lopadova/claude-mem-sync

# Install the plugin (this also installs the mem-sync CLI globally)
claude plugin install claude-mem-sync@claude-mem-sync

The setup hook will automatically:

1. Build the project (if needed)
2. Install the mem-sync CLI globally via npm link
3. Skip installation if mem-sync is already available

Note: The setup hook runs via node (see hooks.json), so you must have Node.js ≥ 18 installed and on your PATH, even if you primarily use Bun. Without Node ≥ 18, the automatic CLI installation will fail. Verify everything is working:

# Check the plugin is active
claude plugin list
# You should see claude-mem-sync@claude-mem-sync with status: ✔ enabled

# Check the CLI is available
mem-sync --help

Local development: If you cloned the repo locally, you can add it as a local marketplace instead:

claude plugin marketplace add /path/to/claude-mem-sync
claude plugin install claude-mem-sync@claude-mem-sync

Already installed the CLI manually? No problem — the setup hook detects existing installations and skips automatically.

What does the hook do? Every time Claude reads a memory (via the mcp__plugin_claude-mem_mcp-search__* tools), the hook records which observations were accessed. This data is used to compute more accurate eviction scores — memories that are actually used by Claude get higher scores and survive longer.

# Option A: Install from GitHub (Bun)
bun add -g @lopadova/claude-mem-sync

# Option B: Install from GitHub (npm — requires GitHub Packages auth)
# 1. Create a PAT at github.com/settings/tokens with read:packages scope
# 2. Configure registry:
echo "@lopadova:registry=https://npm.pkg.github.com" >> ~/.npmrc
echo "//npm.pkg.github.com/:_authToken=YOUR_TOKEN" >> ~/.npmrc
# 3. Install:
npm install -g @lopadova/claude-mem-sync

Step 3: Run the Setup Wizard

The interactive wizard creates your configuration file at ~/.claude-mem-sync/config.json.

mem-sync init

The wizard will ask you:

1. Your developer name — a unique identifier (e.g., alice, bob). This appears in contribution file paths and exported JSON.

2. claude-mem database path — press Enter to accept the default (~/.claude-mem/claude-mem.db), or provide a custom path if you installed claude-mem elsewhere.

3. Project configuration — for each project you want to sync:

   • Project name — a short identifier (e.g., my-app, backend-api)
   • Remote repo — the GitHub repo in owner/name format (e.g., my-org/dev-memories)
   • Auto-merge — yes to push directly, no to create Pull Requests for review
   • Export filters — which observation types/keywords/tags to export

Here’s what a typical config looks like after the wizard:

{
  "global": {
    "devName": "alice",
    "claudeMemDbPath": "~/.claude-mem/claude-mem.db",
    "evictionStrategy": "passive",
    "evictionKeepTagged": ["#keep"],
    "mergeCapPerProject": 500,
    "exportSchedule": "friday:16:00",
    "logLevel": "info",
    "profiles": { "enabled": false, "anonymizeOthers": true },
    "distillation": { "enabled": false, "allowExternalApi": false }
  },
  "projects": {
    "my-app": {
      "enabled": true,
      "remote": {
        "type": "github",
        "repo": "my-org/dev-memories",
        "branch": "main",
        "autoMerge": true
      },
      "export": {
        "types": ["decision", "bugfix", "feature", "discovery"],
        "keywords": ["architecture", "breaking"],
        "tags": ["#shared"]
      }
    }
  }
}

Tip: You can edit ~/.claude-mem-sync/config.json manually at any time to change settings.

Step 4: Create the Shared Team Repository

You need a private repository where all team members push their memory exports.

Recommended: Use setup-repo

mem-sync setup-repo my-team-memories

This interactive wizard will:

• Create the directory structure (contributions/, merged/, profiles/, distilled/)
• Copy CI workflow templates for your provider (GitHub Actions, GitLab CI, or Bitbucket Pipelines)
• Configure the LLM provider for distillation (GitHub Copilot or Anthropic)
• Optionally create the GitHub repo via gh repo create
• Generate a README and .gitignore
• Make the initial commit

Manual alternative

If you prefer to set things up manually:

# Create a new private repo
gh repo create my-org/dev-memories --private --description "Shared AI memories"
gh repo clone my-org/dev-memories
cd dev-memories

# Create directories (works on all platforms — Windows, macOS, Linux)
node -e "['contributions','merged','profiles','distilled'].forEach(d=>{require('fs').mkdirSync(d,{recursive:true});require('fs').writeFileSync(d+'/.gitkeep','')})"

Important: This repo should be private. Observations can contain code snippets, internal URLs, and technical decisions you don’t want public.

Step 5: Set Up GitHub Actions (CI/CD)

If you used mem-sync setup-repo in Step 4, workflows are already in place — skip to Step 6.

GitHub Actions will automatically merge contribution files when developers push exports.

5.1 — Add the Merge Workflow

This workflow triggers every time a developer pushes a contribution file. It merges all contributions, deduplicates, applies eviction caps, and generates developer profiles.

# Make sure you're in the shared repo directory
cd dev-memories

# Create the workflows directory
node -e "require('fs').mkdirSync('.github/workflows',{recursive:true})"

# Download the merge workflow
curl -sL https://raw.githubusercontent.com/lopadova/claude-mem-sync/main/templates/github-action/merge-memories.yml -o .github/workflows/merge-memories.yml

# Download the gitignore
curl -sL https://raw.githubusercontent.com/lopadova/claude-mem-sync/main/templates/.gitignore.example -o .gitignore

The merge workflow will:

1. Detect new files in contributions/
2. Run mem-sync ci-merge to merge + dedup + cap at 500 observations
3. Generate developer profiles in profiles/
4. Commit and push the merged result

5.2 — Add the Distillation Workflow (optional)

If you want LLM-powered knowledge distillation (extracting rules and knowledge docs from merged observations), add the distillation workflow:

curl -sL https://raw.githubusercontent.com/lopadova/claude-mem-sync/main/templates/github-action/distill-knowledge.yml -o .github/workflows/distill-knowledge.yml

Choose your LLM provider:

5.3 — Commit and Push

git add -A
git commit -m "chore: initial setup for team memory sharing"
git push

5.4 — Invite Your Team

Add team members as collaborators:

gh repo edit my-org/dev-memories --add-collaborator teammate-username

Each team member needs to:

1. Install the plugin: claude /install-plugin lopadova/claude-mem-sync
2. Run mem-sync init — using the same repo but their own devName

Step 6: Your First Export

Now let’s export your memories to the shared repo.

6.1 — Preview What Would Be Exported

Always preview first to make sure your filters are working correctly:

mem-sync preview --project my-app

This shows you:

• How many observations match your filters
• The type, date, and title of each matched observation
• The total export size

If nothing matches, check your filter config in ~/.claude-mem-sync/config.json. Make sure export.types includes the types you want (e.g., ["decision", "bugfix"]).

6.2 — Export

mem-sync export --project my-app

This will:

1. Query your local claude-mem database
2. Filter observations based on your config
3. Clone the shared repo
4. Write a JSON file to contributions/my-app/alice/2026-03-19T18-30-00.json
5. Git commit and push

6.3 — Verify on GitHub

Go to https://github.com/my-org/dev-memories and check:

• Your contribution file should appear in contributions/my-app/your-name/
• The GitHub Action should trigger automatically
• After the Action runs, merged/my-app/latest.json should be updated

Step 7: Import Team Memories

Once other team members have exported their memories and the merge bot has processed them, you can import the merged result into your local database.

# Import for a specific project
mem-sync import --project my-app

# Or import all enabled projects at once
mem-sync import --all

This will:

1. Clone the shared repo
2. Read merged/my-app/latest.json
3. Deduplicate against your existing observations
4. Insert new observations into your local claude-mem database

After importing, Claude will have access to your teammates’ discoveries, decisions, and bugfixes during your sessions.

Step 8: Enable Developer Profiles

Developer profiles analyze each team member’s contributions and compute metrics like knowledge spectrum, concept coverage, and contribution quality. No LLM needed — fully deterministic, zero cost.

8.1 — Enable in Config

Open ~/.claude-mem-sync/config.json and set:

{
  "global": {
    "profiles": {
      "enabled": true,
      "anonymizeOthers": true
    }
  }
}

8.2 — Generate Profiles Locally

# Preview what profiles would be generated
mem-sync profile --project my-app --dry-run

# Generate all profiles
mem-sync profile --project my-app

# Generate just your profile
mem-sync profile --project my-app --dev alice

# Also generate markdown (human-readable)
mem-sync profile --project my-app --format md

This creates files in profiles/my-app/:

profiles/
  my-app/
    alice/
      profile.json       # Your full profile data
      profile.md          # Human-readable version
    bob/
      profile.json
    team-overview.json    # Team aggregate stats

8.3 — What’s in a Profile

Each profile contains:

• Knowledge Spectrum — what types of observations you create (e.g., 40% decisions, 30% bugfixes) vs team average
• Concept Map — which concepts you cover and which ones you haven’t touched that the team has (knowledge gaps)
• File Coverage — which directories and files you’ve worked on, plus a specialization index
• Temporal Pattern — how consistently you contribute (observations per week, consistency score)
• Survival Rate — what percentage of your exported observations survived the merge process (quality proxy)

Note: Profiles are automatically generated in CI after each merge if you followed Step 5. You can also run them locally anytime.

Step 9: Enable Knowledge Distillation (optional)

Knowledge distillation uses Claude (via the Anthropic API) to analyze your team’s merged observations and extract:

• Rules — actionable CLAUDE.md-compatible rules with confidence scores
• Knowledge Base — grouped knowledge documentation by concept clusters

Cost: approximately $0.33 per run for 500 observations. See the cost estimation table.

9.1 — Get an Anthropic API Key

1. Go to console.anthropic.com
2. Create an account (or log in)
3. Go to API Keys and create a new key
4. Copy the key (it starts with sk-ant-)

9.2 — Enable in Config

Open ~/.claude-mem-sync/config.json and set:

{
  "global": {
    "distillation": {
      "enabled": true,
      "allowExternalApi": true,
      "model": "claude-sonnet-4-20250514",
      "minObservations": 20
    }
  }
}

Both enabled and allowExternalApi must be true. This double opt-in is intentional — it ensures you consciously decide to send observation data to the Anthropic API.

9.3 — Run a Dry Run First

# Set your API key (or pass it with --api-key)
export ANTHROPIC_API_KEY=sk-ant-your-key-here

# Preview without making an API call
mem-sync distill --project my-app --dry-run

This shows you:

• How many observations would be sent
• Estimated token count
• Estimated cost

9.4 — Run Distillation

mem-sync distill --project my-app

This creates files in distilled/my-app/:

distilled/
  my-app/
    rules.md                    # CLAUDE.md-compatible rules
    knowledge-base.md           # Grouped knowledge documentation
    distillation-report.json    # Run metadata (tokens, cost, stats)
    feedback.json               # Rule accept/reject tracking

9.5 — Review the Output

Open distilled/my-app/rules.md to see the extracted rules. Each rule has:

• The rule statement (imperative, actionable)
• Rationale (why this rule exists)
• Confidence score (50%–100%)
• Evidence count and type breakdown

Rules are suggestions — they are never auto-merged into your CLAUDE.md. Review them, and copy the ones you agree with into your project’s CLAUDE.md manually.

9.6 — Automate in CI (optional)

If you added the distillation workflow in Step 5.2, it runs automatically after each merge and creates a Pull Request with the distilled output. This PR needs human review before merging.

Step 10: Launch the Dashboard

The web dashboard gives you a visual overview of everything: observations, profiles, team insights, distilled knowledge, and more.

mem-sync dashboard

Open your browser at http://localhost:3737.

The dashboard has 9 tabs:

To use a custom port:

mem-sync dashboard --port 8080

Step 11: Set Up Automatic Scheduling (optional)

Instead of running mem-sync export and mem-sync import manually, you can schedule them to run automatically.

Automatic (all platforms)

# Detect your OS and install scheduled tasks
mem-sync schedule install

# To remove them later
mem-sync schedule remove

Manual: Linux (cron)

# Export every Friday at 16:00
0 16 * * 5 mem-sync export --all >> ~/.claude-mem-sync/logs/export.log 2>&1

# Import every Saturday at 09:00
0 9 * * 6 mem-sync import --all >> ~/.claude-mem-sync/logs/import.log 2>&1

# Monthly maintenance (1st of month at 03:00)
0 3 1 * * mem-sync maintain >> ~/.claude-mem-sync/logs/maintain.log 2>&1

Manual: macOS (launchd)

mem-sync schedule install creates plist files in ~/Library/LaunchAgents/.

Manual: Windows (Task Scheduler)

schtasks /create /tn "claude-mem-sync-export" /tr "mem-sync export --all" /sc weekly /d FRI /st 16:00 /rl LIMITED /f
schtasks /create /tn "claude-mem-sync-import" /tr "mem-sync import --all" /sc weekly /d SAT /st 09:00 /rl LIMITED /f
schtasks /create /tn "claude-mem-sync-maintain" /tr "mem-sync maintain" /sc monthly /d 1 /st 03:00 /rl LIMITED /f

Configuration Reference

Config file location: ~/.claude-mem-sync/config.json

Created by mem-sync init or manually. See templates/config.example.json for a full example.

Global Settings

Per-Project Settings

Filter Logic

Filters are combined with OR. An observation is exported if it matches any criterion:

exported = matchesType(obs, types) OR matchesKeyword(obs, keywords) OR matchesTag(obs, tags)

If all filter arrays are empty, nothing is exported (safe default — you won’t accidentally leak data).

Tag System

claude-mem has no native tag system. Tags like #shared and #keep work via free-text search across the title, narrative, and text fields of observations.

• #shared — mark observations for team export
• #keep — protect observations from eviction (score = Infinity)

CLI Reference

Web Dashboard

Launch a local web dashboard to visualize your team’s shared memories, access patterns, profiles, and distilled knowledge.

mem-sync dashboard              # http://localhost:3737
mem-sync dashboard --port 8080  # custom port

Screenshots

Tabs

API Endpoints

Design

• Dark theme with glassmorphism cards and gradient accents
• Chart.js 4 for interactive visualizations
• Inter font, animated counters, hover effects
• Responsive layout (sidebar collapses on mobile)
• Zero frameworks — vanilla JS SPA, single HTML file
• Reads directly from local SQLite databases + contribution/profile/distillation files

Developer Knowledge Profiles

Generate per-developer analytics from contribution and merged data — no LLM required, zero API cost, fully deterministic.

What It Computes

Each developer profile contains 5 metrics:

Team Overview

When profiles are generated for multiple developers, a team-overview.json is also produced with:

• Total developers count
• Average observations per developer
• Average survival rate across the team
• Average concept diversity (coverage percentage)
• Aggregated type distribution

Team Concepts & Knowledge Gaps

The team concepts analysis identifies knowledge bus-factor risks — concepts known by only 1 developer. Visible in the dashboard’s Team Insights tab and available via the /api/team/concepts endpoint.

Privacy Model

• Profiles are opt-in (enabled: false default)
• No developer rankings or cross-dev comparisons
• When anonymizeOthers is enabled, all comparisons use anonymized team averages
• A developer controls their own visibility by enabling/disabling export

Knowledge Distillation

Analyze merged team observations with an LLM to extract actionable rules and knowledge documentation. Produces CLAUDE.md-compatible rules and grouped knowledge patterns.

What It Produces

How It Works

1. Loads merged observations from merged/{project}/latest.json
2. Filters out excluded types (configurable via distillation.excludeTypes)
3. Builds a structured prompt with system instructions and observation data
4. Calls the Anthropic Messages API (Claude Sonnet 4 by default)
5. Parses the JSON response using Zod schema validation
6. Writes rules.md, knowledge-base.md, distillation-report.json, and feedback.json

Cost Estimation

Costs based on Claude Sonnet 4 pricing ($3/MTok input, $15/MTok output).

Rules Quality

Each distilled rule includes:

• Confidence score (0.5–1.0) — based on evidence count and developer diversity
• Source count — number of observations supporting the rule
• Source types — which observation types contributed evidence
• Dev diversity — how many different developers contributed supporting observations
• Category — architecture, testing, security, performance, conventions, workflow, data, or dependencies

Rules below 0.5 confidence are not included. Rules are suggestions requiring human review — they are never auto-merged into CLAUDE.md.

Feedback Loop

The dashboard’s Distilled tab provides accept/reject/modify buttons for each rule. Feedback is stored in distilled/{project}/feedback.json and can be incorporated into future distillation runs (the next run can exclude rejected rules).

Privacy Safeguards

• allowExternalApi: false by default — must be explicitly enabled
• Observations are pre-filtered before API calls (optional excludeTypes)
• System prompt instructs the LLM: no specific code snippets, file paths, or developer names in output
• Provenance cites counts and types only, not attribution
• Output delivered as PR, never auto-merged

Eviction & Scoring

The Problem

Without eviction, databases grow unboundedly. A team of 12 developers sharing weekly would accumulate thousands of observations per project within months.

Hook Mode (recommended)

Uses real access data from the PostToolUse hook:

score = (type_weight * 0.3) + (recency_weight * 0.2) + (access_weight * 0.5)

Passive Mode (fallback)

No hook required. Uses diffusion across developers as a value proxy:

score = (type_weight * 0.4) + (recency_weight * 0.3) + (diffusion_weight * 0.3)

diffusion_weight = devs_who_have_it / total_devs

#keep — Protecting Critical Memories

Observations with #keep in their title, narrative, or text get score = Infinity and are never pruned.

Configuring Weights

"eviction": {
  "strategy": "hook",
  "scoring": {
    "typeWeight": 0.3,
    "recencyWeight": 0.2,
    "thirdWeight": 0.5
  }
}

Weights must sum to 1.0.

Destination Patterns

Pattern A: Dedicated repo (recommended)

One repo per team/org for all project memories. Clean separation, one Action to maintain.

Pattern B: In-project folder

Memories stored in .shared-memories/ inside each project repo. Simpler but adds JSON to code repos.

Pattern C: Hybrid

Dedicated repo with a pointer file (.claude-mem-sync.json) in each project repo.

Maintenance

What mem-sync maintain Does

1. Backup — copies DB to claude-mem.db.backup
2. Pruning — removes low-score observations older than threshold (respects #keep)
3. FTS rebuild — rebuilds all FTS5 indexes
4. Optimize — runs ANALYZE and VACUUM
5. Integrity check — if it fails, restores from backup

Emergency: Restoring from Backup

cp ~/.claude-mem/claude-mem.db.backup ~/.claude-mem/claude-mem.db

Security & Privacy

What Gets Exported

Only observations matching your configured filters (types, keywords, tags) are exported. Empty filters = nothing exported.

Best Practices

• Always preview first: mem-sync preview before your first export
• Use PR review mode: set autoMerge: false for human review
• Private repos only: the shared memories repo should always be private
• Review your filters: observations can contain code with secrets, tokens, or internal URLs

Developer Profiles Privacy

• Opt-in: profiles.enabled is false by default
• No rankings: profiles show individual metrics vs anonymized team average — never cross-developer comparisons
• anonymizeOthers: true (default): comparisons use “team average”, never naming other developers
• Self-controlled: a developer controls their visibility by enabling/disabling export

Knowledge Distillation Privacy

• Double opt-in: both distillation.enabled and distillation.allowExternalApi must be explicitly true
• Type exclusion: excludeTypes keeps sensitive observation types out of API payloads
• No code/names in output: system prompt instructs the LLM to never include code snippets, file paths, or developer names
• PR-based delivery: distilled output is delivered as a pull request — never auto-merged
• Provenance by counts: rules cite “3 bugfix observations” not “from Alice’s session”

Architecture

• Runtime — works on Bun (v1.0+)
• SQLite — bun:sqlite on Bun
• Export + hook are read-only on claude-mem’s DB
• Import is the only write operation — uses transactions with rollback safety
• access.db is a separate tracking database — never touches claude-mem’s schema
• PRAGMA busy_timeout = 5000 on all connections for WAL contention handling
• Multi-provider git — GitHub, GitLab, Bitbucket with optional self-hosted host override
• Array-based process spawning — all shell commands use child_process.spawn with array args (no shell injection)
• Profiler — reads contribution/merged JSON files, computes metrics deterministically (no LLM, no API)
• Distiller — direct fetch to Anthropic Messages API (no SDK dependency), Zod-validated structured output
• Dashboard — pure Node.js HTTP server, 19 API endpoints, vanilla JS SPA with Chart.js 4

Publishing a Release (maintainers)

This package is published automatically to GitHub Packages every time a GitHub Release is created. No custom tokens are needed — the workflow uses the built-in GITHUB_TOKEN.

How it works

1. A maintainer bumps the version
2. The release script commits, tags, pushes, and creates a GitHub Release
3. The release-package workflow triggers automatically
4. It installs dependencies, runs tests, type-checks, builds, and publishes to GitHub Packages
5. The package becomes available at @lopadova/claude-mem-sync

Automated release (recommended)

The release script updates the version in all files (package.json, plugin.json, .claude-plugin/plugin.json, .claude-plugin/marketplace.json), commits, tags, pushes, and creates a GitHub Release — all in one step.

# Interactive — prompts for major/minor/patch
bun run release

# Or pass the bump type directly
bun run release patch   # bug fixes, small tweaks
bun run release minor   # new features, backward compatible
bun run release major   # breaking changes

Manual release

If you prefer to do it step by step:

# 1. Note the current version
cat package.json | grep '"version"'

# 2. Update the version in ALL these files:
#    - package.json
#    - plugin.json
#    - .claude-plugin/plugin.json
#    - .claude-plugin/marketplace.json (under plugins[0].version)

# 3. Commit
git add package.json plugin.json .claude-plugin/plugin.json .claude-plugin/marketplace.json
git commit -m "Bump version from X.X.X to Y.Y.Y"

# 4. Tag
git tag vY.Y.Y

# 5. Push commit and tag
git push && git push --tags

# 6. Create the GitHub Release (triggers the publish workflow)
gh release create vY.Y.Y --title "vY.Y.Y" --generate-notes

The workflow runs at .github/workflows/release-package.yml.

“Config not found”

Run mem-sync init to create the config file.

“CLI is required for PR-based export”

Install the CLI for your provider:

• GitHub: gh
• GitLab: glab
• Bitbucket: curl (uses REST API)

Or set autoMerge: true to use direct push instead of PR/MR mode.

“No observations match export filters”

Check your filter config. Use mem-sync preview to see what matches. All empty filters = nothing exported (safe default).

Hook not tracking accesses

Verify the plugin is installed: claude /plugin list. The hook matches tools prefixed with mcp__plugin_claude-mem_mcp-search__.

“Distillation disabled” or “External API disabled”

Both distillation.enabled and distillation.allowExternalApi must be true in your config. This double opt-in is intentional for privacy.

Profiles show “No contributions found”

Make sure the contributions/ directory exists and contains exported JSON files. Run mem-sync export first, or check that the shared repo has been cloned to the current directory.

License

MIT

Made with ❤️ in Florence