Support shared memory across multiple projects (not just global or per-project)

A UserPromptSubmit hook can inject shared context from a common location into any project:

SHARED_DIR="$HOME/.claude/shared-memory"
[ -d "$SHARED_DIR" ] || exit 0
CONTEXT=""
for f in "$SHARED_DIR"/*.md; do
    [ -f "$f" ] || continue
    CONTEXT+="$(cat "$f")
"
done
[ -z "$CONTEXT" ] && exit 0
jq -n --arg ctx "$CONTEXT" '{
  "hookSpecificOutput": {
    "hookEventName": "UserPromptSubmit",
    "additionalContext": $ctx
  }
}'
exit 0

Add to global ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [{
      "matcher": "",
      "hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/shared-memory-loader.sh" }]
    }]
  }
}

Then organize shared memory by topic:

~/.claude/shared-memory/
├── team-conventions.md      # coding standards shared across all repos
├── infra-patterns.md        # deployment/infrastructure knowledge
├── api-configs.md           # shared API endpoints and auth patterns
└── org-context.md           # organizational decisions and context

For subset sharing (e.g., only frontend projects), add a project filter:

if [ -f "./package.json" ]; then
    for f in "$HOME/.claude/shared-memory/frontend"/*.md; do
        [ -f "$f" ] && CONTEXT+="$(cat "$f")
"
    done
fi

This gives you the missing middle layer: global memory for everything, shared-memory directories for project groups, and project-specific memory for individual repos.

Built a plugin for this: claude-memory-bridge

It adds a namespace layer between global and project memory at ~/.claude/shared-memory/<namespace>/. Projects subscribe to namespaces, so you can share conventions across related projects without polluting global scope.

Works with Claude Code CLI and Cowork desktop. MCP server over stdio, no database — reads/writes Claude's native memory format directly.

Would love feedback if anyone tries it out! Issues and suggestions welcome 🙏

The problem is very real, and your solution is great for high level users. That said, if a non-programmer is using regular Claude AI, it seems a "simple" tagging feature would be also useful. Many projects have nothing to do with coding.

still relevant in the current version (2.1.191)

still relevant in the current version (2.1.220)

A current workaround is to give a selected group of repositories the same explicit memory scope instead of copying files between Claude auto-memory directories.

I maintain GoodMemory. With v0.7.1, install the Claude Code integration once, then enable each related repository with the same workspaceId:

npm install -g goodmemory@0.7.1

# Run setup from the first repository.
goodmemory setup --host claude

goodmemory enable claude --workspace-root ~/work/frontend --workspace-id sediwork
goodmemory enable claude --workspace-root ~/work/backend  --workspace-id sediwork
goodmemory enable claude --workspace-root ~/work/infra    --workspace-id sediwork

goodmemory status claude --workspace-root ~/work/frontend
goodmemory status claude --workspace-root ~/work/backend

Those repositories use the same exact workspace scope in the installed local SQLite store, while unrelated repositories keep their own derived workspace IDs. The integration supplies scoped recall through Claude hooks plus read-only MCP trace/stats/inspection; durable writeback remains opt-in and can be reviewed or disabled.

Important boundary: this does not merge or modify Claude Code's native ~/.claude/projects/<path>/memory/ directories, and it does not implement the proposed native shared-memory configuration. It is a parallel local memory scope that can cover the practical multi-repo case today. Manual CLAUDE.md policy and repository instructions should still remain the explicit source of truth.

Disclosure: @hjqcan maintains GoodMemory. This comment was drafted with Codex assistance under @hjqcan's authorization; the commands and scope behavior were checked against the current v0.7.1 CLI and public repository.