auto-memory

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

🆕 What's New in v0.5.0

Codex CLI support (trial) — session-recall-codex reads OpenAI Codex CLI's local SQLite session store (~/.codex/) with a fixed-schema pre-flight before every query. Ships with the base install — no extra, no env var. Full setup →

Also recent (v0.4.0): Claude Code support — session-recall-cc reads Claude Code JSONL sessions and builds an FTS5 index for structured recall (pip install auto-memory[claude]). Full changelog →

Zero-dependency CLI that turns Copilot CLI's local SQLite into instant recall — no MCP server, no hooks, read-only, schema-checked. ~50 tokens per prompt.

Works With

Claude Code user? pip install auto-memory[claude] && export SESSION_RECALL_ENABLE_CLAUDE_BACKEND=1 — then ask your agent to read deploy/install-claude-code.md for full integration.

Codex CLI user? pip install auto-memory && session-recall-codex schema-check — then ask your agent to read deploy/install-codex.md to wire recall into ~/.codex/AGENTS.md.

VS Code user? Ask your agent to read deploy/install-other-backends.md — it walks through detection, env var setup, and verification automatically.

Coming soon: Cursor

Quickstart

pip install auto-memory           # or: git clone + ./install.sh
pip install --upgrade auto-memory # upgrade to latest
session-recall health          # verify it works

Now give your agent a memory. See the 🤖 Agentic Install section below for agent-driven setup, or follow deploy/install.md manually.

📋 Example: Remember Your Last 10 Sessions

Want to recall the last 10 sessions you worked on, which project folder, branch, and what you did? Add this prompt to your agent instructions:

Use `session-recall list --json --limit 10` to show my last 10 sessions.
Display: date, time, full session_id, summary, branch, turns count, project folder.
Format as a table.

Your agent runs it and gets:

 #  │ Date  │ Time  │ Session ID                                 │ Summary                    │ Branch            │ Project                  │ Turns
────┼───────┼───────┼────────────────────────────────────────────┼────────────────────────────┼───────────────────┼──────────────────────────┼──────
 1  │ 05-01 │ 14:32 │ a1b2c3d4-e5f6-7890-abcd-ef1234567890       │ Refactor Auth Module       │ main              │ ~/Projects/acme-api      │   12
 2  │ 05-01 │ 11:15 │ b2c3d4e5-f6a7-8901-bcde-f12345678901       │ Fix CI Pipeline            │ feat/oauth2       │ ~/Projects/acme-api      │   28
 3  │ 04-30 │ 22:48 │ c3d4e5f6-a7b8-9012-cdef-123456789012       │ Add Redis Caching          │ feat/cache-layer  │ ~/Projects/acme-api      │   35
 4  │ 04-30 │ 16:05 │ d4e5f6a7-b8c9-0123-defa-234567890123       │ Debug WebSocket Timeout    │ main              │ ~/Projects/realtime-chat │    8
 5  │ 04-29 │ 09:33 │ e5f6a7b8-c9d0-1234-efab-345678901234       │ Write Migration Script     │ feat/db-v2        │ ~/Projects/realtime-chat │   42
 6  │ 04-29 │ 08:12 │ f6a7b8c9-d0e1-2345-fabc-456789012345       │ Update API Docs            │ main              │ ~/Projects/acme-api      │    3
 7  │ 04-28 │ 19:55 │ a7b8c9d0-e1f2-3456-abcd-567890123456       │ Deploy Staging             │ release/v2.1      │ ~/Projects/acme-api      │   15
 8  │ 04-28 │ 14:20 │ b8c9d0e1-f2a3-4567-bcde-678901234567       │ Code Review Fixes          │ feat/oauth2       │ ~/Projects/acme-api      │   22
 9  │ 04-27 │ 21:07 │ c9d0e1f2-a3b4-5678-cdef-789012345678       │ Scaffold React Components  │ feat/dashboard-ui │ ~/Projects/acme-frontend │   31
 10 │ 04-27 │ 10:44 │ d0e1f2a3-b4c5-6789-defa-890123456789       │ Setup Terraform Infra      │ infra/aws-prod    │ ~/Projects/acme-infra    │   19

~50 tokens. Your agent gets full context of what you worked on — no re-explaining, no wasted searches.

🤖 Agentic Install (for AI coding tools)

If you are an AI coding agent reading this README, follow this protocol to install auto-memory autonomously:

1. Read deploy/install.md — the agent-runnable install guide. It auto-detects the user's environment, prompts for confirmation before any write, handles fresh/upgrade/skip branches, and chains into backend setup.
2. If the user uses Claude Code, read deploy/install-claude-code.md — pip extra, env var, CLAUDE.md wiring, and optional scheduler setup. Ask before any write.
3. If the user uses Codex CLI, read deploy/install-codex.md — no extra or env var needed; wires ~/.codex/AGENTS.md with confirmation. Trial build: schema-check / list / repos.
4. If the user opts in to VS Code / JetBrains / Neovim backends, read deploy/install-other-backends.md (Section 8 of install.md will prompt for this).
5. All files have YAML front-matter declaring requires-user-confirmation: true and mutates-shell-rc / mutates-agent-instructions. Honor these flags. Do not write silently.

💡 Model recommendation: Use a reasoning model (Sonnet 4.6, GPT-5.4, or equivalent) for install execution. Mini/fast models may hallucinate during multi-step install flows.

Quick start for agents (one-liner):

Read deploy/install.md and execute it for the user. Ask before any disk write.

Files an agent should read in order:

🪟 Windows (WSL2) — Enable the Session Store. (IF MISSING)

On Windows 11 + WSL2 (Suggested method of using on Windows install WSL2), IF Copilot CLI doesn't create session-store.db
You must enable it first — otherwise you'll get error: database not found.

Missing Session Store Error Workaround— Enable the Session Store.

Inside any Copilot CLI session, run:

/experimental

Select SESSION_STORE to enable it. This turns on the SQLite-based session store for cross-session history, file tracking, and search.

Verify it worked:

ls ~/.copilot/session-store.db    # should exist after your next session

Note: This is a one-time setup. Once enabled, the session store persists across all future sessions.

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:

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

Before & After

Before auto-memory — 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 auto-memory — same scenario:

You: Fix the failing test in the auth module

Agent: [auto-recall: auto-memory 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: auto-memory 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

Mental Model: RAM vs Disk

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

auto-memory 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

┌─────────────────────────────────────────────────┐
│  copilot-instructions.md                        │
│  "Run auto-memory FIRST on every prompt"         │
└──────────────────┬──────────────────────────────┘
                   │ agent reads instruction
                   ▼
┌─────────────────────────────────────────────────┐
│  auto-memory CLI                                │
│  (pure Python, zero deps, read-only)            │
└──────────────────┬──────────────────────────────┘
                   │ SELECT ... FROM sessions
                   ▼
┌─────────────────────────────────────────────────┐
│  ~/.copilot/session-store.db                    │
│  (SQLite + FTS5, owned by Copilot CLI binary)   │
└─────────────────────────────────────────────────┘

• Zero dependencies — stdlib only (sqlite3, json, argparse)
• Read-only — never writes to ~/.copilot/session-store.db
• 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

Works with Claude Code (opt-in)

In addition to GitHub Copilot CLI, auto-memory can read Claude Code session logs from ~/.claude/projects/. This ships as a separate CLI binary (session-recall-cc) that is off unless explicitly enabled — Copilot CLI users pay zero cost.

Quick install (3 steps)

pip install auto-memory[claude]
export SESSION_RECALL_ENABLE_CLAUDE_BACKEND=1
session-recall-cc health

Optional: pre-warm scheduler

Skip this unless you want consistent sub-100ms latency. The CLI already does on-demand indexing (~50ms after the first run).

If you have the repo cloned, the installer auto-detects your OS:

bash scripts/install-claude-sidecar.sh --install    # auto-detects OS
bash scripts/install-claude-sidecar.sh --status     # check state
bash scripts/install-claude-sidecar.sh --uninstall  # remove

Otherwise, pick your OS below:

macOS (Intel & Apple Silicon)

Uses launchd — Apple's native scheduler. Survives sleep/wake, auto-runs at login.

PYTHON3="$(command -v python3)"
cat > ~/Library/LaunchAgents/com.session-recall-cc.sidecar.plist <<EOF
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
  "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0"><dict>
  <key>Label</key><string>com.session-recall-cc.sidecar</string>
  <key>ProgramArguments</key><array>
    <string>$PYTHON3</string><string>-m</string>
    <string>session_recall.providers.claude_code.sidecar</string>
    <string>--once</string>
  </array>
  <key>EnvironmentVariables</key><dict>
    <key>SESSION_RECALL_ENABLE_CLAUDE_BACKEND</key><string>1</string>
    <key>PATH</key><string>/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin</string>
  </dict>
  <key>StartInterval</key><integer>300</integer>
  <key>RunAtLoad</key><true/>
  <key>StandardOutPath</key><string>$HOME/.claude/.sr-sidecar.log</string>
  <key>StandardErrorPath</key><string>$HOME/.claude/.sr-sidecar.err.log</string>
</dict></plist>
EOF
launchctl load -w ~/Library/LaunchAgents/com.session-recall-cc.sidecar.plist

Verify: launchctl list | grep session-recall-cc Uninstall: launchctl unload -w ~/Library/LaunchAgents/com.session-recall-cc.sidecar.plist && rm ~/Library/LaunchAgents/com.session-recall-cc.sidecar.plist

Linux / Windows WSL2

Uses cron — works on any distro and WSL2.

(crontab -l 2>/dev/null; echo "*/5 * * * * env SESSION_RECALL_ENABLE_CLAUDE_BACKEND=1 $(command -v python3) -m session_recall.providers.claude_code.sidecar --once >> ~/.claude/.sr-sidecar.log 2>&1") | crontab -

Verify: crontab -l | grep session-recall Uninstall: crontab -l | grep -v session-recall | crontab -

[!TIP] Full details (Windows Task Scheduler, interval tuning, troubleshooting) → deploy/install-claude-code.md §7

How it works

• Reads Claude Code's per-conversation ~/.claude/projects/<cwd>/<uuid>.jsonl files
• Builds an FTS5 index at ~/.claude/.sr-index.db
• Indexer is mtime-incremental — typical update <100ms
• Read-only — never modifies Claude Code's session files

Sample output

Try: "What work was done on the API integration in this repo? Check session-recall-cc for recent history."

session-recall-cc list --limit 5

session-recall-cc search "database migration" --limit 3

🤖 For your AI coding agent

If you'd rather have an agent install this for you, point it at deploy/install-claude-code.md. The doc has YAML front-matter (requires-user-confirmation: true) and per-step "agent: ask user" prompts so a reasoning model walks you through install with confirmation at every mutating step.

Works with Codex CLI (trial, opt-in)

auto-memory can also read OpenAI Codex CLI's local SQLite session store (~/.codex/). This ships as a separate binary (session-recall-codex) that is inert unless invoked — Copilot CLI and Claude Code users pay zero cost, and a Codex storage change can never break the main CLI.

Quick install (2 steps)

pip install auto-memory          # binary included, no extra needed
session-recall-codex schema-check   # validates ~/.codex storage before any query

Then wire it into Codex (with confirmation) via deploy/install-codex.md, which appends the recall block from codex-instructions-template.md to ~/.codex/AGENTS.md — Codex reads that file every session, so the block fires only when Codex is the invoker.

Trial build scope

How it works

• Reads ~/.codex/state_5.sqlite + ~/.codex/thread_history_1.sqlite, read-only (mode=ro + PRAGMA query_only)
• Every data command runs a fixed-schema pre-flight first — if a Codex upgrade changes the storage schema, the CLI refuses cleanly (exit 2/4) instead of returning wrong data
• Sub-agent and archived threads are excluded by default (higher-level context only)
• Verified read-only: the bundled 15-test smoke set (codex-test/codex-test-set.md) proves byte-identical store hashes before/after

Try (in Codex): "What did I work on in this repo recently? Run session-recall-codex list --json --limit 5."

Usage

Try these prompts with your agent

Once wired into your agent's instruction file, session-recall 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.

session-recall files --json --limit 10
session-recall list --json --limit 5

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

session-recall search "specific term" --json

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

session-recall show <session-id> --json

Operational commands:

session-recall health          # 9-dimension health dashboard
session-recall schema-check    # validate DB schema after Copilot CLI upgrades
session-recall repos --json     # discovered repositories across providers

session-recall now supports multi-storage discovery for current Copilot CLI layouts, including session-state sources when legacy session-store.db is absent.

Multi-Storage Recall

By default, session-recall reads only from Copilot CLI's SQLite database. Enable file-backed providers to also recall sessions from VS Code, JetBrains, and Neovim:

export SESSION_RECALL_ENABLE_FILE_BACKENDS=1

Supported Providers

Lookback Defaults

Trust Model

File-backed content is marked _trust_level: "untrusted_third_party" and wrapped in sentinel fences (<<UNTRUSTED-FILE-BACKED-CONTENT>>). CLI content is _trust_level: "trusted_first_party". Downstream agents can use these markers to fence untrusted content.

Environment Variables

Health Check

session-recall health runs a multi-dimensional diagnostic. The core 9 dimensions check your Copilot CLI SQLite database:

session-recall health

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)

Per-Provider Health

Use --provider to diagnose a specific backend. Each provider gets 4 sub-dimensions:

session-recall health --provider cli

 10 Provider:cli
      ├─ Path Discovery     🟢 GREEN  10.0  2 path(s) found
      ├─ File Inventory     🟢 GREEN  10.0  DB 13656064 bytes, 457 state file(s)
      ├─ Recent Activity    🟢 GREEN  10.0  20 session(s) in last 30d
      └─ Trust Model        🟢 GREEN  10.0  Trusted first-party; fences disabled

session-recall health --provider vscode

 10 Provider:vscode
      ├─ Path Discovery     🟢 GREEN  10.0  1 path(s) found
      ├─ File Inventory     🟢 GREEN  10.0  12 file(s)
      ├─ Recent Activity    🔴 RED     0.0  No sessions in last 5d
      └─ Trust Model        🟢 GREEN  10.0  Untrusted third-party; output fences enabled

Available providers: cli, vscode, jetbrains, neovim, all

JSON output for agent parsing:

session-recall health --provider vscode --json
# Returns structured JSON with providers.vscode.dimensions[] array

Not enabled? If you see error: vscode provider is not enabled, set the env var first:

export SESSION_RECALL_ENABLE_FILE_BACKENDS=1
# Or follow the full guide: deploy/install-other-backends.md

Agent Integration

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

See deploy/install.md for setup and copilot-instructions-template.md for integration patterns.

See UPGRADE-COPILOT-CLI.md for schema validation after Copilot CLI upgrades.

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. auto-memory is strictly read-only. It never writes to ~/.copilot/session-store.db.

What happens when Copilot CLI updates its schema? Run session-recall schema-check to validate. The tool fails fast on schema drift rather than returning bad data. See UPGRADE-COPILOT-CLI.md.

Roadmap

See ROADMAP.md.

Contributing

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

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

🔗 Share it: "Zero-dependency CLI that gives your AI coding agent session memory. Read-only, schema-checked, ~50 tokens per prompt." → github.com/dezgit2025/auto-memory

Disclaimer

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

Contributors

• @jshessen — Multi-storage provider architecture (PR #5)

License

MIT