Lore
Open-source memory infrastructure

One memory layer,
shared across all your agents.

Your agents keep forgetting. Lore gives them a durable, shared memory graph — so knowledge from Claude Code, Codex, OpenCode and your custom agents lives in one place, inspectable and under your control.

Works with your agents

Claude CodeCodexopencodeOpenClawHermes AgentPiModelContextProtocol

The problem

Every agent starts from zero. Every time.

You tell your agent the same preferences, fix the same mistakes, re-explain the same project context — session after session, agent after agent.

Without shared memory

  • Agent forgets your preferences next session
  • Two agents repeat each other’s mistakes
  • Project knowledge locked in one chat history
  • Switching runtimes means starting over
  • No way to see what the agent “remembered”

With Lore

  • Every agent boots the same stable baseline
  • Knowledge written once, recalled by all
  • Memories have addresses — preferences://user
  • One token connects Claude Code, Codex, your stack
  • Inspect every recall before the agent trusts it

Multi-agent memory

A shared memory graph, not isolated silos.

Write a preference once. Every agent you run reads the same durable graph. Memory becomes infrastructure your agents share, not a private scratchpad each one rewrites.

Claude CodeClaude CodeCodexCodexopencodeOpenCodeOpenClawOpenClawHermes AgentHermesPiPi

Each memory lives at a stable URI. Agents read and write through the same paths — no copy-paste, no drift, no duplicated knowledge.

Use cases

Where shared memory pays off.

Multi-agent collaboration

Three agents, one source of truth.

A planning agent captures project decisions. A coding agent reads them before every change. A review agent checks them against new code. No one re-explains the architecture — it’s already in the graph.

core://agent + project:// → shared
Shared memory graph across agents

Cross-runtime

Switch tools, keep the memory.

You wrote half the project with Claude Code, now you’re in Codex. Both read the same ~/.lore baseline. Your preferences, conventions, and past decisions travel with you — not with the editor.

one token · all runtimes
Recall workbench — inspectable

Trust & inspection

See exactly what the agent remembered.

Before every reply, Lore injects a small recall set. You can inspect which memories were considered, which were selected, and why. No black-box memory deciding things for you.

read-before-trust · rollback history
Recall analysis — what was used

How it works

Three things happen every turn.

01 / Boot

Load the baseline

At session start, agents read stable identity, preferences, and project memory from core://, preferences://, project://.

02 / Recall

Inject before reply

Before each answer, a small <recall> set is injected — only the memories relevant to this prompt.

03 / Inspect

Read before trust

Agents open a node before relying on it. You see the full recall trace and can roll anything back.

Open source

Self-host it. Or let us run it.

Lore is MIT-licensed and runs on your machine with Docker. Prefer zero ops? Use the managed cloud and connect agents with a token.

One command to self-host

Spins up Postgres + Redis + web, connects your agent runtimes, writes config. Re-run anytime to update.

$npx @loremem/cli
$docker pull fffattiger/lore

MIT

License, fully self-hostable

GitHub →
MIT licenseDocker pullsLatest release

FAQ

Questions before you start.

Stop re-explaining. Start compounding.

Free to self-host. Free tier on the cloud. 5 minutes to connect your first agent.