Lore
中文Start free

Agents forget,
Lore remembers

Long-term memory for Claude Code, Codex, OpenCode and every agent you run. Your rules load at session start, and the right memories arrive before every reply.

Start freeSelf-host
npx @loremem/cli

Works natively with

  • Claude CodeClaude Code
  • CodexCodex
  • opencodeOpenCode
  • PiPi
  • OpenClawOpenClaw
  • Hermes AgentHermes
  • ZCode
  • DeepSeekDeepSeek Harness
Claude Code — checkout-apiLore connected
Session
Memory graphrecall ≥ 0.60 · max 3

Every agent keeps its own instruction file. Copy a rule into five of them and it stops being one rule. Lore keeps each memory once, with an address and a trigger, and hands every agent only what the moment needs.

How it works

Four steps around every session.

Hooks and plugins run them for you. The agent needs no extra instructions to remember.

01Boot

Same start, every session

A fixed baseline loads first: workflow rules, persona, your preferences and notes for this runtime.

02Recall

Only what clears the bar

Before each reply, Lore scores memories against the prompt. At most three get in, and anything under 0.60 is dropped.

03Read

Open before trusting

Recall hands over addresses. The agent opens the node it needs, and Lore records which recalls were actually used.

04Write back

Lessons outlive the session

What should last goes back into the graph. Every change lands in history with a diff.

If Lore is unreachable, plugins fail open and agents keep working.

Memory

A memory is a node, not a chat log.

Each memory has an address, a priority, a trigger for when it applies, and a body you can read like a spec.

Address
A stable URI. Paths form a tree, and Dream keeps them tidy.
Priority
P0 holds at most 5 memories and P1 at most 15. The rest compete on relevance.
Disclosure
Says when the memory should surface, so recall can find it at the right moment.
History
Every edit is kept with a diff, whoever made it.
project://checkout_api/webhooksP1
Disclosure

When changing refund or payment webhooks in checkout-api

Glossary

refund webhookidempotency keystripe

Content

Refund webhooks are idempotent by event.id. Tests must replay the same event twice and assert a single refund row.

History
  1. rev 4Dream sharpened the trigger
  2. rev 3Claude Code added the replay rule
  3. rev 1Created from Codex

Dream

Memory that tidies itself. Every step can be undone.

Dream runs at night. It merges duplicates, sharpens vague triggers and moves drifting paths back. Everything lands in a diary you review, and the latest run can be rolled back.

Dream diary

  • Merged 2 duplicates intoproject://checkout_api/testing

  • Sharpened the trigger onpreferences://user/writing

  • Moved a drifting path toproject://checkout_api/webhooks

Roll back this run
Dream
03:00

every night

Backups
04:00

keeps 7, disk or WebDAV

Recall floor
0.60

lower scores are dropped

Per prompt
3

memories at most

SkillsLoremem Cloud

Skills, versioned and loaded on demand.

Write a skill once as a SKILL.md package. Agents see the list at session start and load the one a task needs, or you call it by name.

  1. 1

    Listed at session start

    Agents see names and descriptions. Nothing downloads yet.

  2. 2

    Loaded when the task fits

    lore_skill_get fetches the package into ~/.lore/skills.

  3. 3

    Or run it by name

    Type $release-notes and Lore loads it for that request.

  • release-notes

    Write release notes from merged pull requests

    v3
  • stripe-replay

    Replay Stripe events against the sandbox

    v2
  • db-migrate

    Plan and run schema migrations safely

    v5

›$release-notes for 2.4

⏺ lore_skill_get(release-notes)

⎿ Loaded release-notes v3

Agents

One memory. Every agent.

Install once with npx @loremem/cli. It finds the runtimes on your machine and wires each one in natively.

  • Claude CodeClaude CodePlugin · hooks + MCP
  • CodexCodexPlugin · hooks + MCP
  • opencodeOpenCodeNative plugin · no MCP
  • PiPiExtension
  • OpenClawOpenClawPlugin
  • Hermes AgentHermesMemory provider
  • ZCodePlugin · MCP bridge
  • DeepSeekDeepSeek Harness
  • ModelContextProtocolAny MCP clientMemory tools only, no automatic recall

Deploy

Start in the cloud, or own every byte.

Loremem Cloud

Free to start
  • Sign in with GitHub or Google
  • Models preconfigured
  • Versioned skills
  • Organizations and projects
Start free

Self-hosted

MIT · open source
  • One command deploys Lore, PostgreSQL and Redis
  • Bring any OpenAI-compatible model
  • Daily backups to disk or WebDAV
  • Skills are Cloud-only
Self-hosting guide

FAQ

Questions, answered.

What happens when Lore is down?

Plugins fail open. Agents keep working, without recall, until Lore is back.

Do I need to change how I prompt?

No. Boot and recall run through hooks and plugins, so you work the way you already do.

Which memories load at session start?

A fixed baseline: core://agent, core://soul, preferences://user and core://agent/<runtime>. Everything else is recalled per prompt.

Can I undo something Dream changed?

Yes. The latest Dream run can be rolled back, and every node keeps its history with diffs.

Does it work with Chinese?

Yes. Search uses the pg_jieba tokenizer, and the console is bilingual.

Give your agents a memory.

Create a token, run one command, and every agent you use reads and writes the same graph.

Start freeRead the docs