| name | memory-compression |
| description | claude-mem plugin — automatic cross-session memory compression and injection via SQLite + semantic search. Captures tool usage, compresses observations, and injects relevant context at session start. |
| tags | ["skill","memory","plugin","claude-mem"] |
| triggers | ["compress memory","search past sessions","what did we do last time","memory is getting bloated","retrieve campaign history"] |
Memory Compression (claude-mem)
What It Is
claude-mem is a Claude Code plugin (v11.0.0) installed at ~/.claude/plugins/marketplaces/thedotmack/. It provides automatic cross-session memory that is distinct from our manual brain/memory/ file system.
- Storage: SQLite at
~/.claude-mem/claude-mem.db (auto-created on first use)
- Vector search: Chroma embeddings at
~/.claude-mem/chroma/ (semantic similarity)
- Viewer UI:
http://localhost:37777 (available when worker is running)
- Runtime: Requires Bun (auto-installed on first session start via
smart-install.js)
How It Coexists With Our Memory System
| Layer | System | What It Stores | Managed By |
|---|
| brain/ + memory/ | Bravo's file-based memory | Business context, tasks, decisions, patterns | Bravo (manual writes) |
@modelcontextprotocol/server-memory | MCP knowledge graph | Named entities and relations | Bravo via MCP tools |
| claude-mem SQLite | Plugin automated memory | Tool usage observations, session summaries | claude-mem hooks (automatic) |
These three layers are complementary, not conflicting. claude-mem captures low-level tool activity automatically — it does not replace the structured business memory in brain/ and memory/.
Activation — Fully Automatic
5 lifecycle hooks run without any action from Bravo or CC:
| Hook | Trigger | What It Does |
|---|
Setup | Session with "startup/clear/compact" | Runs smart-install.js — installs Bun/uv if missing, starts worker |
SessionStart | Session with "startup/clear/compact" | Starts worker service, injects relevant past context |
UserPromptSubmit | Every prompt | Initializes session-level observation tracking |
PostToolUse | Every tool call | Records observation (file edits, bash commands, etc.) |
Stop | End of response | Compresses session into summary observation |
SessionEnd | Session end | Marks session complete in DB |
Privacy: Wrap sensitive content in <private>...</private> tags to prevent storage.
Available Skills (Auto-Loaded by Plugin)
| Skill | Trigger | Purpose |
|---|
/mem-search | Use when asking "did we already solve this?" | 3-layer search: index → timeline → fetch |
/make-plan | Plan a multi-phase feature | Orchestrator + subagent phased planning |
/do | Execute a plan from /make-plan | Subagent-driven plan execution |
/smart-explore | Explore codebase structure | Guided codebase discovery |
/timeline-report | Generate activity timeline | Chronological work history |
Search Workflow (When CC Asks About Past Work)
Always use the 3-layer approach — never fetch full observations without filtering first (10x token savings):
Step 1 — search(query="...", limit=20, project="Business-Empire-Agent")
Step 2 — timeline(anchor=<id>, depth_before=3, depth_after=3)
Step 3 — get_observations(ids=[...]) # Only for IDs that look relevant
Worker Service
The worker (Express on port 37777) handles AI processing asynchronously. It requires Bun.
npx claude-mem start
npx claude-mem status
npx claude-mem stop
open http://localhost:37777
The worker auto-starts when a session begins with "startup", "clear", or "compact" in the first prompt. For other sessions, tool hooks still record observations — but search requires the worker to be running.
Interaction With Our Existing Memory System
claude-mem reads, Bravo writes. claude-mem passively records what Bravo does. Bravo continues writing to brain/STATE.md, memory/ACTIVE_TASKS.md, etc. as per Rule 0.
No conflicts with MCP memory server. @modelcontextprotocol/server-memory stores named entities (knowledge graph). claude-mem stores time-series observations (what happened, when, what files). Different data models, different use cases.
When to use claude-mem search vs. reading memory files:
- Reading
memory/SESSION_LOG.md → current session's structured decisions and task status
/mem-search → tool-level observations from any past session (what files were touched, what commands ran)
Configuration
Settings auto-created at ~/.claude-mem/settings.json on first run. No manual configuration required for basic operation.
Installation Record
- Installed: 2026-04-06
- Version: 11.0.0
- Method:
npx claude-mem@11.0.0 install --ide claude-code
- Plugin dir:
~/.claude/plugins/marketplaces/thedotmack/
- Registered in:
~/.claude/settings.json under enabledPlugins
- Bun status: Not yet installed — auto-installs on first session with "startup" prompt
Obsidian Links
- [[skills/memory-management/SKILL]] | [[brain/CAPABILITIES]] | [[memory/SESSION_LOG]]
Maven-specific adaptation
Maven compresses long campaign histories into summarized retros (brain/retros/) once a campaign has run >30 days and stabilized. Active campaign data (last 14 days) stays full-fidelity in data/pulse/ and memory/SESSION_LOG.md. Compression preserves: thesis, audience, creative angle, spend, results, and the one-line lesson. It discards: per-day metric noise, transient creative variants, exploration dead-ends.
