nem

nem

Your agent forgets. nem doesn't.

nem versions agent context the way git versions code. When an LLM compacts its context window, it discards reasoning, resolved edge-cases, and prior decisions. nem persists that context explicitly, makes it searchable and portable across agents and machines — and the agent uses it with the same commands a human would.

Features

• Ingests Codex and Claude Code sessions (conversation + reasoning + tool calls, with tool outputs truncated so they don't burn tokens).
• Immutable commits: each commit copies the text of the range (a snapshot), not a pointer — what you saved never changes.
• Full-text search with BM25 ranking (SQLite FTS5), prioritizing conversation over tool noise.
• Git-based sync with automatic secret redaction: API keys and tokens (OpenAI, Anthropic, HuggingFace, AWS, GitHub, GitLab, Stripe, wandb, …), connection strings, Authorization headers, and sensitive env vars are masked before anything leaves your machine.
• Agent skill: nem init installs a SKILL.md into Claude Code and Codex so the agent knows when and how to use nem on its own.
• Single binary, SQLite embedded in pure Go (no cgo) → a static executable. Works offline.

Stack

Go · GORM over glebarez/sqlite (modernc, no cgo) · SQLite FTS5/BM25 · cobra · git (via os/exec).

Install

Windows (Scoop):

scoop bucket add Dieg0Code https://github.com/Dieg0Code/scoop-bucket
scoop install nem

macOS / Linux (Homebrew):

brew install Dieg0Code/homebrew-tap/nem

With Go:

go install github.com/Dieg0Code/nem/cmd/nem@latest

Prebuilt binaries: download the archive for your OS/arch from the Releases page and put nem on your PATH.

From source:

go build -o nem ./cmd/nem

Usage (MVP)

nem init                          # creates ~/.nem + git, installs the agent skill
nem ingest                        # ingest Codex and Claude (or: nem ingest codex)
nem status                        # active session + uncommitted messages

nem add -L 20                     # stage the last 20 messages
nem add --from <msgID> --to <id>  # ...or an exact range
nem commit -m "decision about X"  # immutable snapshot

nem log                           # commit history
nem read HEAD --format llm        # the commit, clean for an agent
nem search "decay" --format llm   # BM25 search (--role all includes tools)

nem remote add origin <url>       # configure the remote
nem sync                          # export (redacting) → git push → import
nem clone <url>                   # clone the store on another machine

Output formats

read and search accept --format:

• llm — clean, no metadata, for agent ingestion
• json — structured
• markdown — human-readable (default)

The agent skill

nem init installs a SKILL.md into the agents present on your machine (~/.claude/skills/nem/ and ~/.codex/skills/nem/). It teaches the agent to recall prior context at the start of a session (nem status / nem search / nem read HEAD) and to persist resolved decisions (nem add + nem commit, writing its own commit message). nem sync stays in your hands.

nem init --no-skill    # skip skill installation
nem skill install      # (re)install the skill later, e.g. after upgrading nem

nem only ever owns the nem skill directory — it never touches your other skills.

Security: secret redaction

Scrubbing runs only on nem sync (the one boundary where content leaves your machine). The local DB keeps the raw text — just like ~/.codex and ~/.claude already do — so ingesting adds no new exposure. nem sync redacts and reports:

$ nem sync
exported 12 commits
redacted 7 secrets: 5 huggingface-token, 2 env-secret
synced with the remote

Architecture

cmd/nem/            entrypoint
internal/
  cli/              cobra commands
  config/           paths (~/.nem, nem.db, store/)
  db/               GORM models + Store (interface) + FTS5
  ingest/           codex/claude parsers (Parser interface)
  session/          active-session detection
  output/           snapshots + llm/json/markdown formats
  redact/           secret detection and masking
  skill/            embedded SKILL.md + installer
  sync/             JSONL export/import + git

Every component follows the interface + functional options pattern: New*(...) (Interface, error) returns the interface, which makes testing with mocks trivial.

Development

go test ./...        # full suite
go vet ./...

The race detector (go test -race) requires cgo. On Windows run the tests from PowerShell (not Git Bash, which DLL-shadows the mingw toolchain). In CI it runs on Linux (.github/workflows/ci.yml).