traceary

README

Traceary

日本語

Traceary is a local-first CLI and MCP server for recording and searching AI agent work logs, session boundaries, and shell command audits.

Install the CLI first, then add the plugin for your AI agent host to enable automatic recording.

Why Traceary

AI-assisted development gets messy quickly when:

• session context disappears after clear or compact
• Git history explains what changed, but not always why
• shell command output is hard to connect back to the right agent or session
• work is split across Claude, Codex, Gemini, and manual terminal steps
• multiple sessions and worktree moves make the timeline harder to follow

Traceary keeps those records in one local SQLite store so the same history can be reused from the CLI, hooks, and MCP clients.

Three-layer model

Traceary is no longer just a local event log. v0.5.0 organizes the product around three layers that map to how agent workflows actually need context.

Layer	What lives there	How it is fed
Audit / Archive	raw events (prompts, transcripts, command audits), session boundaries	host hooks (SessionStart, UserPromptSubmit / BeforeAgent, PostToolUse / AfterTool, Stop / AfterAgent, PreCompact / PreCompress, SessionEnd) — see host coverage matrix
Working memory	handoff / context packs assembled from recent sessions	derived on demand by traceary session handoff / MCP get_context. Claude PreCompact digest also syncs into sessions.summary so timeline / handoff have a useful header before the session ends
Durable memory	reusable facts such as decisions, constraints, preferences, and artifact refs	curated through the traceary-memory-review skill (review-intent triggers) and the traceary-memory-remember skill (explicit-write triggers)

In practice, Traceary acts as a local-first memory substrate for AI agents: hooks feed L1 mechanically, L2 is recomputed when the next session starts, and L3 stays small because it only grows when the operator (or an explicit "remember that" verb) says so.

Traceary is local-first. It writes to SQLite on your machine and does not include built-in telemetry, analytics, or hosted storage.

Getting started

Step 1: Install the Traceary CLI

The CLI is required first — agent plugins invoke the traceary binary via hooks.

# Homebrew (recommended)
brew tap duck8823/traceary https://github.com/duck8823/traceary
brew install traceary

# or go install
go install github.com/duck8823/traceary@latest

Tagged releases also publish macOS and Linux archives on GitHub Releases. See the release guide for packaging details.

Step 2: Install the plugin for your agent host

Claude Code (guide)

/plugin marketplace add duck8823/traceary
/plugin install traceary

Codex (guide)

git clone https://github.com/duck8823/traceary ~/src/traceary
cd ~/src/traceary
codex   # then, inside Codex: /plugins -> Traceary Plugins -> Traceary

The traceary integration codex install helper was retired in v0.14.0 and the cleanup-only traceary integration codex uninstall surface was removed in v0.15.0. Use Codex CLI's official /plugins flow shown above for install / uninstall. See the Codex plugin guide for migration details and manual cleanup steps for legacy installs.

Gemini CLI (guide)

bash <(curl -sL https://raw.githubusercontent.com/duck8823/traceary/main/scripts/install-gemini-extension.sh)

For the integration overview, use the native integrations guide. Direct Anthropic API users can also try the experimental native memory-tool backend.

Step 3: Verify

traceary doctor

Quick start

traceary store init is optional. Commands create the database and run migrations on demand. Use store init only when you want to create the DB path up front or confirm write permissions before a session starts. The v0.8.x top-level alias traceary init was removed in v0.14.0; running it now exits with a usage error pointing at traceary store init. See the CLI stability and deprecation policy for the full list of v0.14 removals and replacements.

1. Start a session and write a note

sid=$(traceary session start --client dogfood --agent codex)
event_id=$(traceary log --client dogfood --agent codex --session-id "$sid" --id-only "Investigating failing tests")
traceary show "$event_id" --json

2. Record command output in the same session

traceary audit \
  --client dogfood \
  --agent codex \
  --session-id "$sid" \
  --command "go test ./..." \
  --input '{"stdin":""}' \
  --output '{"stdout":"panic: boom","stderr":"stacktrace","exitCode":1}'

traceary search boom --json
traceary session active

3. Use script-friendly output when needed

traceary log --id-only "Investigating failing tests"
traceary audit --id-only --command "go test ./..." --input '{}' --output '{}'
traceary session end --session-id "$sid" --id-only

4. Promote reusable context when it matters

traceary memory store remember \
  --type decision \
  --workspace github.com/duck8823/traceary \
  --fact "Use traceary session handoff --compact-only for compact resume context" \
  --evidence issue:#502

traceary session handoff --workspace github.com/duck8823/traceary

5. Curate the durable-memory inbox

traceary memory ... is grouped by intent: memory inbox for candidate review, memory store for deliberate writes (remember / propose / distill), and memory admin for extraction, host-side I/O (import / export / activate), maintenance (hygiene / graph), and lifecycle (supersede / expire / set-validity). Daily-read commands (memory search / memory show / memory list) stay top-level. The flat verbs from earlier releases (memory remember, memory accept, ...) were removed in v0.15.0 after the v0.14 compatibility window; use the canonical grouped paths above.

For interactive review of the candidate inbox at the terminal:

traceary memory inbox review
traceary memory inbox review --workspace github.com/duck8823/traceary --type preference --limit 10

memory inbox review is a TTY-only walk-through built on a shared Bubble Tea TUI. Inside the screen a accept, x reject, s skip, e edit/distill, v view evidence, ? help, q quit. Edit/distill never auto-accepts LLM-authored text — it routes through traceary memory store distill and requires you to type the operator-authored fact. Non-interactive shells receive a refusal with exit code 2 and a pointer to the script-friendly fallback (memory inbox list plus memory inbox accept|reject, both of which now also accept a positional id and --id-only).

Inspect recent and live activity

Traceary ships complementary inspection views so you can switch between "what's happening now" and "what happened across a span" without leaving the terminal:

When	Command	Use it to
Watching the workspace dashboard	traceary top	browse active sessions, recent failures / commands, candidate memories, and stale memories in one TUI
Following what is happening now	traceary tail	confirm hooks are firing, watch failures in real time
Understanding what happened across a span	traceary timeline	see gap-separated work blocks with a per-workspace activity summary
Inspecting raw events directly	traceary list / traceary search	jump to an exact kind / session / query
Resuming with assembled working memory	traceary session handoff	start a follow-up session with curated context

traceary top

traceary top

top opens a five-pane Bubble Tea dashboard for active sessions, recent failures, recent commands, candidate memories, and stale memories. Use tab / shift+tab to move between panes, / to filter the focused pane incrementally, and Enter to drill into the highlighted session, event, or memory detail. In non-TTY shells, traceary top --snapshot and traceary top --snapshot --json expose the same data for scripts, including the stale_memories envelope key.

traceary tail

$ traceary tail --limit 3
07:06:44  command_executed  sess=4a70c526  ws=traceary  ls ~/.traceary 2>&1; find ~ -name "traceary…
07:06:47  command_executed  sess=4a70c526  ws=traceary  ./traceary timeline --db-path /Users/duck88…
07:06:52  command_executed  sess=4a70c526  ws=traceary  timeout 1 ./traceary tail --db-path /Users/…

Compact single-line rows (local time by default) fit inside ~100 columns. Add --wide --utc to restore the pre-v0.6.1 tab-separated seven-column layout byte-for-byte, or --json for NDJSON when piping into tooling.

traceary timeline

$ traceary timeline --limit 2
2026-04-15 06:37 - 07:06 (29m21s) total events: 165
  github.com/duck8823/traceary (153) — 自律的に進めてください。
  github.com/duck8823/dotfiles  ( 12) — rust インストールしました
2026-04-15 05:39 - 06:10 (31m1s) total events: 136
  github.com/duck8823/traceary (136) — <analysis> This conversation is a resumption after compaction. …

Each block shows one sub-row per workspace with an activity summary picked from the fallback chain compact_summary → first prompt → kind counts, so you can see what was being done in each workspace instead of just a comma-joined list. Pass --utc to switch text timestamps to UTC; --json adds a workspace_breakdown array alongside the existing block fields.

See docs/cli/README.md for the full flag reference and more examples.

Host capture matrix

The query surface is shared: once Traceary is installed, every host can use the same CLI and MCP memory/context commands. What differs is how much context each host can capture automatically via hooks.

Host	Session lifecycle	Tool audit	Prompt capture	Compact-summary capture	Automatic capture tier
Claude Code	Full	Bash + MCP + failure hooks	Yes	Yes	Full
Codex	Full (SessionStart + Stop)	Tool hooks	Yes	No	Partial
Gemini CLI	Full (SessionStart + SessionEnd)	Tool hooks	No	No	Basic

2026 Q2 note: Claude Code's SubagentStop / PreCompact hooks and Gemini CLI 0.38.x's memory-manager preview are available but not wired into Traceary's managed hook set. The Codex memory feature flag in ~/.codex/config.toml changes Codex's own capture behaviour, not Traceary's — traceary memory admin import codex works regardless. traceary doctor surfaces the same notes under <client>-host-capabilities, and the full list lives in the hook contract.

For the full contract and hook semantics, see the hook contract and event lifecycle.

Defaults worth knowing

• traceary log and traceary audit reuse the latest non-stale active session for the resolved workspace when --session-id is omitted; when remote.origin.url is missing inside a git worktree, Traceary falls back to the worktree root path
• traceary session active treats sessions older than 24h as stale unless you pass --allow-stale
• traceary session start prints a session ID; traceary session end prints the recorded event ID
• traceary session list --json includes label, summary, and parent_session_id when present
• default operator-facing CLI output is English; set TRACEARY_LANG=ja when you want Japanese messaging
• --json output stays language-neutral

Documentation

Use the documentation index for the full map. The most common next pages are:

• Native integrations
• Architecture principles
• Durable memory guide
• CLI reference
• CLI stability and deprecation policy
• Hooks guide
• Hook contract and capability tiers
• Event lifecycle
• MCP guide
• Environment and storage notes

Contributing and support

• bug reports and feature requests belong in GitHub Issues
• security reports should follow SECURITY.md
• this is an actively evolving v0.x OSS tool, so check the changelog before upgrading automation around it

Documentation

Overview

Package main defines the Traceary entrypoint. Environment loading and dependency wiring happen only in this package.