Skip to content

What is LightMem2

LightMem2 is an open plugin platform for long-running AI agents. It lets you build agent capabilities once and run them across multiple agent hosts — OpenClaw, Codex, Claude Code, and future hosts.

LightMem2 vs. TokenPilot

A common point of confusion: LightMem2 and TokenPilot are not the same thing.

LightMem2TokenPilot
What it isA plugin platform and runtimeA plugin that runs on LightMem2
RoleLoads, manages, and executes pluginsProvides cache-aware context management
ScopePlatform-wide: plugins, adapters, CLIOne specific capability: reducing token usage
StatusActive developmentStable — the first official plugin

Think of LightMem2 as the operating system and TokenPilot as an app that runs on it.

What Problems It Solves

  • Long sessions get expensive. As agent sessions grow, every turn carries more context, which means more tokens and higher costs.
  • Context is repetitive. Much of what gets sent to the model each turn is identical to the previous turn — wasteful if not cached.
  • Tool output is noisy. Large tool responses can pollute future turns with irrelevant data.
  • Sessions don't prune themselves. Without eviction, old context accumulates until sessions hit limits or become too slow.

LightMem2 addresses these through its plugin model. TokenPilot, the first plugin, implements the runtime policies: stable-prefix rewriting, context reduction, and lifecycle-aware eviction.

What It Doesn't Solve

  • Short, single-turn interactions. If your sessions are always one-shot, there is nothing for caching or eviction to optimize.
  • Model quality or accuracy. LightMem2 doesn't change model behavior — it changes what context gets sent to the model.
  • All memory problems. Long-term memory is an experimental feature area; TokenPilot focuses on the current session's context window.

Relationship to the Paper

The TokenPilot paper describes the cache-efficient context management technique. LightMem2 is the platform that hosts TokenPilot as its first plugin, and will host additional plugins in the future.

Next Steps

Released under the MIT License.