enam
Persistent memory for Claude Code — at ~350 tokens ambient cost, with no Docker, no API keys, and no MCP servers. Engram captures your sessi
Open source Open in the app JSON README (API)
About
Persistent memory for Claude Code — at ~350 tokens ambient cost, with no Docker, no API keys, and no MCP servers. Engram captures your sessions structurally on PreCompact (zero LLM cost), extracts atomic memories via a single lightweight LLM digest (preferences, practices, current context), and detects emergent patterns from your history — files you always edit together, recurring errors, tool usage anomalies — stored as an auditable Obsidian-compatible wiki. Key features: • Per-project handoff notes — at session start you get a narrative message from the last session in this project • Atomic UPSERT memories keyed by topic — latest wins, no drift, no embeddings • Emergent pattern detection (co-edits, error recurrence, project streaks, tool anomalies) • Durable preferences stay global; ephemeral context is scoped per project • 4-flag CLI surface: --stats, --query, --memories, --forget • Everything local: SQLite at ~/.claude/memory.db, markdown wiki at ~/.claude/patterns/
Details
- Kind
- Plugins
- Topic
- AI, RAG & memory
- Publisher
- sebastianbreguel
- Origin
- marketplace
- Category
- ferramentas
- Stars
- 2
- Last push
- 2026-06-08T23:39:51Z
- Repository state
- ativo
- Language
- Python
- License
- MIT
- Added
- 2026-08-30 01:48:58
- Updated
- 2026-08-30 01:48:58
- Origin id
sebastianbreguel/engram/enam
README
# WhereWasI [](https://github.com/sebastianbreguel/wherewasi/actions/workflows/test.yml) [](https://www.python.org/) [](LICENSE) **Claude forgets everything between sessions.** What you were doing, what's next, what broke — gone the moment you close the terminal. WhereWasI fixes that one thing well. Open a project tomorrow and the first thing you see is *the last task you were on and the next step* — so you pick up immediately. Per-project, automatic, no config. This is **not** a memory bank. No search, no recall by topic, no accumulated history. Just the current state, to the point. ## What you see When you open Claude Code (or **Codex CLI**) in a project, WhereWasI injects a short resume at the top of the session: ``` # where was i: wherewasi · branch wherewasi-v2 Last: Rewriting docs for the resume pivot (README + architecture). Next: Run the full test suite + ruff, then deploy with install.sh. 3 uncommitted · a1b2c3d docs rewrite ``` Two load-bearing lines — **Last** (where you were) and **Next** (what to do) — over one compact git line for grounding. No file dumps, no stale error echoes (you have `git` for that). Zero latency: it's read from a file, no LLM call at open. ## How it works WhereWasI is **files + hooks. No database.** One Markdown file per project at `~/.claude/wherewasi/resume/<cwd-slug>.md` (the cwd path, slashes turned to dashes), refreshed as you work: 1. **Every 25 prompts** (`UserPromptSubmit`, **no LLM**) — a cheap refresh of the git line (branch, uncommitted count, last commit). Your `Last`/`Next` narrative is preserved. 2. **On compaction** (`PreCompact`, **LLM**) — `claude --print` reads the transcript tail and rewrites `Last` + `Next` from what you were actually doing. 3. **On session end** (`SessionEnd`, **LLM**) — rewrites `Last` + `Next` when you leave, so a short session that never compacted isn't stale the next time you open. Runs detached, so it never delays your exit. 4. **On session start** (`SessionStart`) — the resume file is read and injected (with git refreshed live). First time in a project? You get a minimal git-derived resume so you never open blank. Replace semantics: the file always reflects the current state. No history accumulates. ## Install **Requirements:** [Claude Code](https://docs.anthropic.com/en/docs/claude-code) and/or [Codex CLI](https://developers.openai.com/codex), `python3` (stdlib only — no `uv`, no pip, no deps). **As a Claude Code plugin (recommended):** ```bash # In Claude Code: /plugin install wherewasi@sebastianbreguel/wherewasi ``` **Or clone and run the installer:** ```bash git clone https://github.com/sebastianbreguel/wherewasi.git cd wherewasi && ./install.sh ``` > Use **the marketplace install OR `./install.sh`, not both** — each registers the hooks separately, so doing both fires every hook twice (and pays the compaction LLM cost twice). **Codex CLI (also supported):** `./install.sh` auto-detects `~/.codex` (or `$CODEX_HOME`) and wires Codex too. Codex hooks share Claude Code's JSON stdin/stdout wire protocol, so the same `wherewasi.py` runs unchanged. It needs `[features].hooks = true` in `~/.codex/config.toml`, and Codex may prompt to trust the hook on first launch. Resume state is **shared** with Claude (one `~/.claude/wherewasi/resume/<slug>.md` per project), so switching between the two on the same repo shows the same "where was I". Note: Codex has no `SessionEnd` event, so the end-of-session LLM rewrite is skipped there — `PreCompact` + the rolling git refresh keep it fresh. > **First session in a project:** you get a minimal git-only resume. > **After you've worked a bit:** the resume fills in with your last task + next step. ```bash # Uninstall (also cleans up legacy engram data, if any) cd wherewasi && ./uninstall.sh ``` ## Commands ```bash python3 ~/.claude/tools/wherewasi.py --cwd "$PWD" # print this project's resume python3 ~/.claude/tools/wherewasi.py --reset --cwd "$PWD" # clear this project's resume ``` Or use the slash command: `/wherewasi` (show) · `/wherewasi --reset` (clear). ## Privacy and transparency Everything lives in `~/.claude/wherewasi/resume/<cwd-slug>.md` (plain Markdown). Nothing leaves your machine except the one LLM call below. - **Stored**: project name, git branch + uncommitted count + last commit, and the two LLM-written `Last`/`Next` lines. - **NOT stored**: no full transcripts, no source code, no secrets from `.env`. - **LLM calls**: on compaction (`PreCompact`) and at session end (`SessionEnd`), `claude --print` (default Sonnet 4.6 — override with `WWI_MODEL`, or set it empty for your account default) reads the **tail of the transcript** to write two lines. It runs under your existing Claude Code login — **no separate API key**. Set `WWI_SKIP_LLM=1` to disable it entirely (the resume still refreshes from git on the rolling path). Note: from **2026-06-15**, `claude -p` on subscription plans (Pro/Max/Team/Enterprise) draws from a monthly [Agent SDK credit](https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan) you claim once. If it runs out, the LLM rewrite pauses but the rolling git refresh and session-start display keep working. ## Environment variables | Variable | Default | Effect | |---|---|---| | `WWI_SHOW_BANNER` | `1` | Set to `0` to suppress the visible banner (context still injects) | | `WWI_SKIP_LLM` | unset | Set to `1` to skip the LLM rewrites (PreCompact + SessionEnd); resume still refreshes from git | | `WWI_MODEL` | `claude-sonnet-4-6` | Model for the LLM rewrite. Set empty to use your account default | | `WWI_DIGEST_EVERY` | `25` | Rolling-refresh cadence (prompts per `UserPromptSubmit` refresh) | ## How it compares | | WhereWasI | claude-mem | OpenMemory | cortex | |---|---|---|---|---| | Ambient token cost | **~120** | ~2K+ | ~1K+ (MCP) | ~3K (27 tools) | | External services | None | Agent SDK worker | Docker + MCP server | MCP server | | API keys required | No | Yes | No | No | | Runtime | Python (stdlib) | Node worker | Docker | Rust binary | | Storage | One MD file/project | SQLite + worker | Docker volume | Rust store | | Install | `./install.sh` | npm + worker | docker compose | cargo | ## Docs - [Architecture](docs/architecture.md) — files + hooks, the resume file, capture flow - [CLI Reference](docs/cli-reference.md) — commands, hooks, environment variables - [Privacy Policy](docs/privacy.md) — what's stored, what's not, network activity ## License MIT