Lethe — Persistent Memory for AI Agents

Lethe is a self-hosted persistent memory platform for AI agents. It gives

agents durable, reviewable memory that survives sessions, restarts, and model

changes — without surrendering that memory to a third party.

Any agent can use Lethe: ChatGPT, Claude, Claude Code, Cursor, VS Code

agents, OpenClaw, custom MCP clients, local IDE assistants, or backend

services. Lethe is a general-purpose memory platform; the

OpenClaw plugin is one excellent integration, not a

requirement.

Why it exists

Every agent session starts from zero. The same question, the same reasoning,

the same dead ends — every time. Lethe breaks that loop with two

complementary memory systems:

• Session memory — events, summaries, checkpoints, and interruption

recovery so an agent resumes exactly where it stopped.

• Versioned memory — durable, shared, review-gated knowledge with

immutable history, provenance, branching, and merging (see

Memory Git).

Who it is for

• Individuals giving a single agent (ChatGPT, Claude, Cursor, an IDE

assistant) a long memory.

• Teams running several agents against one shared, reviewed memory

backend.

• Operators who need memory to stay on their own infrastructure, in one

container, with their keys.

Three runtime modes

Details and decision guidance: docs/runtime-modes.md.

Install

# legacy session memory, port 18483
docker run -d --name lethe \
  -p 127.0.0.1:18483:18483 \
  -v ./lethe-data:/data \
  ghcr.io/openlethe/lethe:latest

Memory Git (git mode) runs from `docker-compose.git.yml`; hybrid mode is one

instance with both surfaces. Full instructions, all compose variations, and

source builds: docs/installation.md ·

docs/docker-compose.md

Documentation

Read in order:

1. Overview — what Lethe is and isn't, principles, ecosystem

2. Installation — Docker, git mode, hybrid, source build

3. Runtime modes — legacy vs git vs hybrid, how to choose

4. Memory Git — versioned memory, refs, changesets, proposals, merges

5. Legacy mode — session memory and the OpenClaw experience

6. Architecture — components, boundary, data model, security

7. Configuration — every environment variable

8. Deployment & operations — production checklist, upgrades, backup

9. Compose variations — legacy, git, hybrid, governed stack

10. HTTP API reference — every route, both surfaces

11. OpenClaw — the plugin integration

12. Integrations — ChatGPT, Claude Code, Cursor, MCP, multi-agent

13. Migration & upgrading — version migrations, legacy ↔ Memory Git

14. FAQ

Protocol and deep-dive references:

memory-git-v1.md (the `memory_git/v1` contract) ·

memory-context-bridge.md (projection & manifests) ·

observability.md (metrics, SLOs, alerts)

Related projects

• **Charon** — the MCP authorization

and governance gateway for Lethe: scoped principals, independent review,

replay-proof protected merges, and a fail-closed audit ledger. Recommended

for shared or multi-agent deployments.

• OpenClaw plugin — optional context engine for OpenClaw, published on

ClawHub as `lethe`. Optional; everything works without it.

Compatibility

License

See LICENSE. Issues and contributions:

CONTRIBUTING.md · Security reports:

SECURITY.md