session-memory

Team collective memory plugin for Claude Code.

Auto-capture tool usage → promote to knowledge cards → share via git.

What is this?

session-memory is a Claude Code plugin that gives your team a shared knowledge base that grows automatically as you work.

• Individual: Like claude-mem, it records tool usage and injects context into new sessions
• Knowledge: Events get promoted to Cards — reusable lessons with context
• Team: Cards live in git, so git push shares knowledge with the entire team

Quick Start

# Install the plugin
/plugin install session-memory

# Or load locally for testing
claude --plugin-dir /path/to/claude-session-memory

That's it. Zero config. The plugin starts capturing automatically.

How It Works

You work normally with Claude Code
        │
        ▼
PostToolUse hook captures events (Edit, Write, Bash)
        │
        ├──→ Micro-inject: editing a file? Related cards injected instantly
        │
        ▼
Events accumulate in .claude/memory/events/ (local, gitignored)
        │
        ▼
Patterns detected → Card auto-created (with secret scanning)
        │
        ├──→ Existing card? Evidence appended + importance bumped
        │
        ▼
Cards saved to .claude/memory/cards/ (git committed, team shared)
        │
        ▼
Next session → relevant cards auto-injected (scored by recency/importance/file overlap)

Key Concepts

Events (temporary, local)

Raw tool usage logs. Stored per-user in JSONL format. Gitignored by default.

Cards (permanent, shared)

Knowledge units with structure: type, importance, files, tags. Markdown with YAML frontmatter. Git committed for team sharing.

Card types (ADR-lite):

Type	Purpose	Example
decision	Why we chose this	"PostgreSQL over MongoDB"
constraint	Can't change this	"API rate limit 100/min"
gotcha	Watch out for this	"req.user can be null"
runbook	How to do this	"DB migration steps"
onboarding	New member guide	"Dev environment setup"

Evidence Accumulation

Cards are living knowledge. When the same pattern recurs, evidence is automatically appended to the matching card, and its importance is bumped. Cards get stronger over time.

Summaries (archived, shared)

Session summaries stored per-session. Useful for timeline view and onboarding.

Features

Feature	Description
Auto-capture	PostToolUse hooks record Edit/Write/Bash events
Smart injection	Multi-signal scoring (git diff + interest + git log) with token budget
Micro-injection	Edit a file → related card TL;DR injected instantly
Auto-promotion	Patterns trigger card creation (3+ edits, error→fix, high importance)
Evidence accumulation	Recurring patterns strengthen existing cards
Secret scanning	Cards scanned for API keys, tokens, DB URLs before creation
Issue link extraction	Branch names and commit messages → card refs (GH-123, JIRA-456)
Onboarding	Role-based card packages for new team members
Web viewer	Static HTML dashboard for cards and timeline

Skills

Skill	Description
/session-memory:status	Show memory status and injection preview
/session-memory:timeline	Chronological session history
/session-memory:search	Search cards (keyword + structured: tag:auth type:gotcha file:src/)
/session-memory:card	Manually create a card (--type, --importance)
/session-memory:prune	Clean up old events, detect duplicate cards
/session-memory:onboard	Role-based onboarding (frontend, backend, devops, etc.)
/session-memory:viewer	Generate static HTML viewer for cards and timeline

Storage Layout

.claude/memory/
├── config.json              # Project settings (optional)
├── events/                  # Local only (gitignored)
│   ├── jay@macbook.jsonl
│   └── .interest.jsonl      # File access signals
├── cards/                   # Team shared (git committed)
│   └── mem_20260309_a1b2c3.md
├── summaries/               # Session archives (git committed)
│   └── 2026-03-09__abc123.md
└── viewer.html              # Generated web viewer (optional)

Configuration

Create .claude/memory/config.json to customize (all optional):

{
  "maxInjectTokens": 2000,
  "maxMicroInjectTokens": 200,
  "maxDaysToKeep": 30,
  "autoPromote": true,
  "autoEvidence": true,
  "secretScanning": true
}

See lib/defaults.json for all available options.

Git Integration

Auto-shared via git: cards/ and summaries/ are committed, events/ are gitignored.

Merge strategy (add to your .gitattributes):

.claude/memory/cards/*.md merge=union
.claude/memory/summaries/*.md merge=union

Pre-commit hook (optional): templates/pre-commit-hook.sh scans cards for secrets before commit.

Dependencies

bash + jq. That's it. No Node.js, no Python, no database.

vs claude-mem

	claude-mem	session-memory
Individual memory	Yes	Yes
Team sharing	No (local SQLite)	Yes (git)
Knowledge cards	No (events only)	Yes (ADR-lite)
Evidence accumulation	No	Yes
Micro-injection	No	Yes
Secret scanning	No	Yes
Token budget control	No	Yes
Issue link extraction	No	Yes
Web viewer	Yes (port 37777)	Yes (static HTML)
Role-based onboarding	No	Yes
Dependencies	Bun, Chroma, SQLite	bash, jq
Daemon process	Yes (port 37777)	No
License	AGPL-3.0	MIT

License

MIT