lore

Lore is a working model for collaborating with coding agents, packaged as persistent memory and a protocol layer for agent harnesses — Claude Code (reference baseline), OpenCode, and Codex CLI.

Each session starts from what previous sessions recorded — conventions, gotchas, architecture notes, in-flight work.

Install

Requires bash, Python 3, and Go (builds the TUI). Go can be skipped for knowledge and memory use alone, but the TUI is required for /coordinate — it allows coordinator agents to manage worker sessions.

git clone git@github.com:anticorrelator/lore.git && cd lore
bash install.sh                # --framework claude-code (default) | opencode | codex

Then, in a project you work on:

cd your-project
lore init      # create the knowledge store — or run /bootstrap in an agent session to seed it

--dry-run previews, --uninstall removes (data preserved). Re-run with a different --framework to switch harnesses.

Use

Most of lore is automatic. Hooks load relevant knowledge, active work, and conversation threads at session start, and queue new insights for capture at session end. Normal coding sessions need no lore-specific commands.

TUI — lore with no arguments. Work items, live agent sessions in embedded terminals (spawn, drive, triage), knowledge browsing, review-finding follow-ups, and the settlement view.

Arc window — in Claude Code, a session working on an arc (/coordinate) shows a one-row summary above the prompt. Press its ▴, or run /arcs, to open a window with the step graph, the Brief and decisions one line each, the arc's documents, and each step's work item. It appears only in projects with a lore store, and only while an arc is active. Restart running sessions after installing to load it.

Review panel — /reviews in Claude Code opens the project's open /pr-review results: pick a review, select and edit its comments and findings, choose Comment / Approve / Request changes, preview the payload, and post it. A review a session makes announces itself. The panel changes reviews only through lore followup edit-review and lore followup post-review.

Driving work — slash commands in an agent session:

/coordinate	drive a feature end-to-end across protocol sessions
/spec → /implement	plan a work item, then execute the plan with worker agents
/work	create, resume, and check status of work items
/remember	capture insights from the current conversation
/pr-review	multi-lens PR review (--self for your own branch)

Ceremonies — periodic maintenance:

/bootstrap	seed a new project's store by exploring the codebase
/retro	score how well the system supported a finished work cycle
/evolve	review and apply protocol suggestions accumulated by retro
/renormalize	prune and rebalance the knowledge store

Configuration — the installer persists framework and role→model settings to ~/.lore/config/. lore config show prints them; lore harness enable|disable toggles integration per harness. To turn lore off for one project in Claude Code, add {"env": {"LORE_AGENT_DISABLED": "1"}} to the project's .claude/settings.local.json, or to .claude/settings.json to share the setting. That stops lore's hooks and the arc window there.

The wider CLI (lore search, lore work, …) is used by agents and scripts.

Data

Everything lore keeps lives per-project under ~/.lore/repos/<host>/<org>/<repo>/: knowledge entries (with provenance), work items and plans, conversational threads, and scorecard telemetry. Logic lives in this repo; data lives in ~/.lore/ and survives reinstalls.