claude-mem 1.0.0

Session recall for Claude Code — zero-dependency CLI that gives Claude instant memory of past sessions

claude-mem

Your AI coding agent has amnesia. Here's the fix.

~1,900 lines of Python. Zero dependencies. Saves you an hour a day.

Built by Desi Villanueva

Zero-dependency CLI that gives Claude Code instant session recall — no MCP server, read-only, schema-checked. ~50 tokens per prompt.

Works with: Claude Code · Cursor · Aider

---

Quickstart

pip install claude-mem
claude-mem cc-index        # build the local session index
claude-mem install-mode --setup   # wire SessionStart hook into ~/.claude/settings.json

That's it. Every new Claude Code conversation gets the last ~50 tokens of session context injected automatically.

---

The Problem

Every AI coding agent ships with a big number on the box. 200K tokens. Sounds massive. Here's what actually happens:

200,000  tokens — context window (theoretical max)
120,000  tokens — effective limit before context rot kicks in (~60%)
 -65,000  tokens — MCP tools
 -10,000  tokens — instruction files
=========
 ~45,000  tokens — what you ACTUALLY have before quality degrades

LLMs don't degrade gracefully — once you cross roughly 60% of the context window, the model starts losing coherence. The industry calls it "lost in the middle": attention goes to the beginning (instructions) and the end (recent turns), but your actual working context in the middle gets progressively fuzzier.

I timed it over a week: 68 minutes per day lost to re-orientation after compactions and new sessions.

It's a death spiral of diminishing context — each compaction leaves the agent slightly dumber, which burns more tokens explaining things, which triggers the next compaction sooner.

The Compaction Tax

Every 20–30 turns, the context warning hits and you get two bad choices: ignore it and watch the agent hallucinate, or run /compact and watch it lobotomize itself into a tidy two-paragraph summary of a 30-minute investigation. Either way you lose five minutes re-narrating your own project back to the agent like it's a new hire. That's not a workflow — that's a hamster wheel.

The 200x ROI

Here's the cost comparison that made me build this:

Operation	Tokens	What you get
grep -r "auth" src/	~5,000-10,000	500 results, mostly irrelevant
find . -name "*.py"	~2,000	Every Python file, no context
Agent re-orientation	~2,000	You re-explaining yesterday
claude-mem files --json --limit 10	~50	Exactly the 10 files you touched yesterday

50 tokens vs 10,000 — a 200x improvement.

Before & After

Before claude-mem — new session on a project:

You: Fix the failing test in the auth module

Agent: Let me explore the project structure...
       $ find . -name "*.py" | head -50           ← 2K tokens
       $ grep -r "test.*auth" tests/              ← 5K tokens
       $ cat tests/test_auth.py                    ← 3K tokens
       $ cat src/auth/handler.py                   ← 2K tokens
       I see several test files. Which test is failing?

You: The one we were working on yesterday, the token refresh edge case

Agent: Let me search for that...
       $ grep -r "refresh" tests/                  ← 4K tokens
       ...

Total: ~16K tokens burned, 8 minutes elapsed, agent still isn't oriented.

After claude-mem — same scenario:

You: Fix the failing test in the auth module

Agent: [auto-recall: claude-mem files --json --limit 10]
       → src/auth/refresh.py, tests/test_refresh_edge_cases.py,
         src/auth/token_store.py (last touched 14h ago)

       [auto-recall: claude-mem list --json --limit 3]
       → Yesterday: "Fixed token refresh race condition, one edge case
         test still failing on expired token + network timeout combo"

       I can see from your last session that test_refresh_edge_cases.py
       has a failing test for the expired token + network timeout case.
       Let me look at that specific test...
       $ cat tests/test_refresh_edge_cases.py      ← 1K tokens (targeted)

Total: ~1.1K tokens, 30 seconds, agent is immediately productive.

How it compares

Approach	Dependencies	Writes to DB	Setup	Agent-native
claude-mem	None (stdlib)	❌ Read-only	pip install	✅ Instruction-file
MCP server	Node.js runtime	Varies	Server config	❌ Protocol layer
Custom hooks	Varies	Often yes	Hook scripts	❌ Event-driven
Manual grep	None	❌	None	❌ Manual

Mental Model: RAM vs Disk

• Context window = RAM. Fast, limited, clears on restart.
• session-store.db = Disk. Persistent, searchable, grows forever.

claude-mem is the page fault handler — it pulls exact facts from disk in ~50 tokens when the agent needs them.

It's not unlimited context. It's unlimited context recall. In practice, same thing.

Design

Claude Code backend:

┌─────────────────────────────────────────────────┐
│  ~/.claude/settings.json                        │
│  SessionStart hook → injects ~50-token context  │
└──────────────────┬──────────────────────────────┘
                   │ hook fires on every new session
                   ▼
┌─────────────────────────────────────────────────┐
│  claude-mem CLI                                 │
│  (pure Python, zero deps)                       │
└──────────────────┬──────────────────────────────┘
                   │ SELECT ... FROM FTS5 index
                   ▼
┌─────────────────────────────────────────────────┐
│  ~/.claude/.sr-index.db                         │
│  (SQLite FTS5, built by claude-mem from         │
│   ~/.claude/projects/ JSONL session files)      │
└─────────────────────────────────────────────────┘

• Zero dependencies — stdlib only (sqlite3, json, argparse)
• Read-only on source data — never writes to Claude Code's JSONL files
• WAL-safe — exponential backoff retry on SQLITE_BUSY (50→150→450ms)
• Schema-aware — validates expected schema on every call, fails fast on drift
• Telemetry — ring buffer of last 100 invocations for concurrency monitoring
• Backend auto-detection — auto-detects ~/.claude/projects/ automatically

Usage

Try these prompts with your agent

Once wired into your agent's instruction file, claude-mem runs on every prompt — giving the agent your recent files and sessions as context before it does anything else.

"Search recent sessions about fixing the db connection bug"
"Check past 5 days sessions for latest plans?"
"Pick up where we left off on the API refactor"
"search recent sessions for last 10 files we modified"
"search sessions for the db migration bug"

No special syntax. The agent reads your session history and gets oriented in seconds instead of minutes.

How it works under the hood

Progressive disclosure — most prompts never get past Tier 1.

Tier 1 — Cheap scan (~50 tokens). Usually enough.

claude-mem files --json --limit 10
claude-mem list --json --limit 5

Tier 2 — Focused recall (~200 tokens). When Tier 1 isn't enough.

claude-mem search "specific term" --json

Tier 3 — Full session detail (~500 tokens). Only when investigating something specific.

claude-mem show <session-id> --json

Operational commands:

claude-mem health          # 9-dimension health dashboard

Claude Code Backend

claude-mem reads ~/.claude/projects/ JSONL session files and builds a local SQLite FTS5 index at ~/.claude/.sr-index.db.

Quick setup (2 steps)

# 1. Build the local session index from ~/.claude/projects/ JSONL files
claude-mem cc-index

# 2. Wire a SessionStart hook into ~/.claude/settings.json
claude-mem install-mode --setup

After step 2, every new Claude Code conversation automatically receives ~50 tokens of recent session context.

Index commands

claude-mem cc-index                 # build / update the index (incremental)
claude-mem cc-index --rebuild       # force a full rebuild from scratch
claude-mem cc-index --status        # show index freshness and session count

The index lives at ~/.claude/.sr-index.db. It is owned and written exclusively by claude-mem — Claude Code's JSONL files are never modified.

Hook installation

claude-mem install-mode             # detect Claude Code surfaces (CLI, VS Code, JetBrains, Desktop)
claude-mem install-mode --setup     # wire SessionStart hook automatically
claude-mem install-mode --dry-run   # preview changes before applying

--setup adds a SessionStart hook entry to ~/.claude/settings.json. The hook runs claude-mem list --json --limit 3 at the start of each conversation and injects the result as context (~50 tokens).

Using the Claude Code backend on query commands

# Auto-detection: uses Claude Code backend if ~/.claude/projects/ exists
claude-mem list --json --limit 10
claude-mem files --days 7
claude-mem search "auth refactor"

# Explicit backend flag
claude-mem --backend claude list --json --limit 10
claude-mem --backend claude files --days 7
claude-mem --backend claude search "auth refactor"
claude-mem --backend claude show SESSION_ID
claude-mem --backend claude health

Backend auto-detection rules:

Condition	Backend selected
~/.claude/projects/ exists	claude
Aider history files found	aider
Cursor workspace DBs found	cursor
None found	Error with setup instructions

Health Check

Dim Name                   Zone     Score  Detail
----------------------------------------------------------------------
 1  DB Freshness           GREEN   8.0  15.8h old
 2  Schema Integrity       GREEN  10.0  All tables/columns OK
 3  Query Latency          GREEN  10.0  1ms
 4  Corpus Size            GREEN  10.0  399 sessions
 5  Summary Coverage       GREEN   7.4  92% (367/399)
 6  Repo Coverage          GREEN  10.0  8 sessions for owner/repo
 7  Concurrency            GREEN  10.0  busy=0.0%, p95=48ms
 8  E2E Probe              GREEN  10.0  list→show OK
 9  Progressive Disclosure  CALIBRATING  —  Collecting baseline (n=42/200)

Agent Integration

claude-mem works with any agent that supports instruction files — Claude Code, Cursor, Aider, Windsurf, and more. Installation wires claude-mem into your agent's instruction file so it runs context recall automatically.

What This Isn't

• Not a vector database — no embeddings, SQLite FTS5 only.
• Not cross-machine sync — local only.
• Not a replacement for project documentation — recalls what you did, not how the system works.

FAQ

Is it safe? Does it modify my session data? No. claude-mem is strictly read-only on your agent's session data. It never writes to Claude Code's JSONL files under ~/.claude/projects/. The only file claude-mem writes is its own index at ~/.claude/.sr-index.db.

Roadmap

See ROADMAP.md.

Contributing

See CONTRIBUTING.md for setup and guidelines. Issues, PRs, and docs improvements are welcome.

If claude-mem saved you time, star the repo — it's the best way to help others find it.

Share it: "Zero-dependency CLI that gives Claude Code session memory. Read-only, schema-checked, ~50 tokens per prompt." → github.com/osamarehman/claude-mem

Disclaimer

This is an independent open-source project. It is not affiliated with, endorsed by, or supported by Anthropic or any other company. There is no official support — use at your own risk. Contributions and issues are welcome on GitHub.

License

MIT