Documentation

The company brain your agents act on.

Overview

Lore reads the places your company’s decisions actually live (Slack, Gmail, Notion, and your Claude Code sessions), clusters the recurring ones by topic, and writes each topic as a SKILL.mdfile in Anthropic’s open Agent Skills format. Any MCP-compatible agent loads those skills and runs them. Every line cites the source that set the rule.

How it works

Four stages, fully automatic once connected:

1. Pull. OAuth into each source. Lore reads from where the work happens.
2. Structure. A bi-temporal knowledge graph of every decision, with when it happened and when Lore learned it, so superseded rules age out.
3. Emit. Each topic becomes a versioned SKILL.md. Content-hash versioning bumps it only when something material changed.
4. Run. Your agents load skills via MCP and act, with human approval on consequential steps until trust is established.

Your sources

Live today: Slack, Google Drive, Gmail, Notion. In progress: GitHub, Linear, PagerDuty, Granola. Each is a single OAuth click during onboarding, no YAML or config. Lore extracts the facts that matter and links them across sources, so a decision mentioned in Slack and confirmed over email becomes one cited rule.

Load skills (MCP)

Lore exposes your skills over MCP. Point any MCP client (Claude Code, Cursor, Codex, or your own agent) at your workspace endpoint with your token:

claude mcp add lore --transport http \
  --url https://jointhelore.com/api/mcp/v1/<workspace> \
  --header "Authorization: Bearer lore_<your-key>"

Then the agent can list skills, load a SKILL.md on demand, and cite the facts behind any claim.

Citations & freshness

Every claim in a skill carries an observation id like [obs:a3f1c2]. Resolve it to read the verbatim Slack message, PR, or email that set the rule. The agent follows the rule with the receipt attached, instead of paraphrasing.

When the underlying work changes, the old fact gets a valid_untilset, the skill re-emits, and the citation points at the new source. Your agents always run the current version, not last quarter’s.

Skill vs memory layer

Memory layers (Mem0, Zep, Letta) store notes and hand them back when the agent asks; the agent reads them and improvises. Lore writes the cited playbook the agent executes, versioned in git, re-emitted when a rule changes. We use a memory layer underneath. We are not one. The artifact your agent runs is the SKILL.md.

Getting access

Self-serve signups are paused while we hand-onboard the first cohort of design partners, free. If your agents are running on tribal knowledge buried in Slack, get on the list and we will reach out within 24 hours.