LETHE

Live Ephemeral Token & History Engine — offload big tool outputs out of your LLM agent's context and recall them on demand, so long tasks cost fewer tokens.

🌍 This README is bilingual. English · Español

🔌 Use it in Claude Code or Codex (save tokens now)

LETHE ships as an MCP server. Two lines and your agent can move big outputs out of its context and recall them on demand — fewer tokens on every long task. / LETHE viene como servidor MCP. Dos líneas y tu agente saca outputs grandes del contexto y los recupera cuando los necesita — menos tokens en cada tarea larga.

Claude Code:

pip install "lethe-llm-context[mcp]"
claude mcp add lethe -- lethe-mcp

Codex: add an MCP block to ~/.codex/config.toml — see integrations/codex/mcp-config.md.

Then drop in the guiding skill so the agent archives on its own: integrations/claude-code/SKILL.md.

Tools exposed: lethe_archive · lethe_recall · lethe_status. Full guide: integrations/claude-code/mcp-config.md.

▶️ See it work (no API key) / Míralo funcionar (sin API key)

python -m lethe.examples.mcp_demo

  LETHE — offload big tool outputs, recall on demand, save tokens

The agent runs 4 commands. Each returns a wall of text:

  → build.log               857 tok in context  —archive→  stub '[paged: build.log | handle=6d48a1c2]'
  → pytest.txt              479 tok in context  —archive→  stub '[paged: pytest.txt | handle=1e56b0f4]'
  → db_dump.json            829 tok in context  —archive→  stub '[paged: db_dump.json | handle=ec02...]'
  → trace.txt               414 tok in context  —archive→  stub '[paged: trace.txt | handle=d888...]'

lethe_status:  4 blocks archived, 2579 tokens moved out of context

30 steps later the agent needs a buried fact. It recalls by keyword:

  lethe_recall("launch_code")  →  found launch_code = 4242 (rehydrated losslessly from the archive)

  Context window cost
    without LETHE :  2579 tok  (everything stays resident)
    with LETHE    :    38 tok  (only tiny stubs remain)
    saved         :  2541 tok  (-99%)

This runs the same logic the MCP tools use — lethe_archive / lethe_recall / lethe_status.

English

What LETHE is (today, honestly)

When an LLM agent runs a long task (tens to hundreds of steps), its context window fills with material that was useful but no longer is: stale tool outputs, files read 30 steps ago, big JSON dumps. That means more tokens per turn, higher cost, and eventually the context ceiling.

LETHE gives the agent an external, lossless scratch store so that heavy content lives outside the context window and only a tiny handle stays inside. Two ways to use it:

1. As an MCP server (the install path above). The agent calls lethe_archive(content) on a big output and keeps only the returned 4–8-char handle; later it calls lethe_recall(handle) or lethe_recall("keywords") to bring the full text back. Nothing is ever deleted — recall is lossless. The guiding skill tells the agent when to do this, so it happens near-automatically. This is provider-agnostic: it works in any MCP host (Claude Code, Codex).

2. As a Python library (in-loop context GC). ContextManager runs an automatic pass inside your agent loop: a heuristic Curator scores each block (recency, whether later blocks cite it, block kind, plus an optional cheap-model relevance call) and an Archivist pages the coldest blocks out to a store — losslessly, leaving stubs — to hold the working set under a token budget. Referenced handles are paged back in on demand.

Be clear about the mechanism. An MCP server cannot silently rewrite the host's context window. LETHE works by giving the agent explicit offload/recall tools plus a skill that makes using them near-automatic — not by magic. That honesty is the point.

✅ What works today vs. 🗺️ what's on the roadmap

The multi-provider, ensemble, and semantic-retrieval design below is the long-term vision, not what's implemented. Here is the honest split:

Each milestone ships as a tagged release — full notes in CHANGELOG.md.

The mental model (OS analogy)

LETHE is designed like an operating system managing virtual memory. This analogy guides the architecture; the ✅/🟡 above says how much of it runs automatically today.

How LETHE differs from agent-memory libraries

Mem0, Zep, Letta and friends persist facts across sessions. LETHE targets the opposite: managing the live, in-session working context of a running loop — deciding what to keep in the window right now. It's complementary to a long-term memory product, not a competitor. Whether that difference is decisive is something the roadmap above still has to prove.

Quickstart (no API key needed)

python -m pytest -q                  # run the full test suite, including the needle test
python -m lethe.examples.fake_loop   # WATCH it work: live view, blocks paging out, budget held

Real Claude demo

$env:ANTHROPIC_API_KEY="sk-..."   # PowerShell
python -m lethe.examples.claude_loop

Design docs

• docs/specs/2026-06-12-lethe-vertical-slice-design.md — approved design of the built slice
• docs/LETHE_engineering_design.md — the full long-term engineering vision (multi-provider, ensemble, embeddings). Read it as the roadmap, not the current state.

License

Released into the public domain under the Unlicense. Free for everyone, anywhere.

Español

Qué es LETHE (hoy, con honestidad)

Cuando un agente LLM ejecuta una tarea larga (decenas o cientos de pasos), su ventana de contexto se llena de material que fue útil pero ya no lo es: resultados de herramientas obsoletos, archivos leídos hace 30 pasos, dumps de JSON enormes. Eso significa más tokens por turno, más costo y, al final, el techo de contexto.

LETHE le da al agente un almacén externo y sin pérdida para que el contenido pesado viva fuera de la ventana de contexto y dentro solo quede un handle diminuto. Dos formas de usarlo:

1. Como servidor MCP (la instalación de arriba). El agente llama lethe_archive(content) sobre un output grande y conserva solo el handle de 4–8 caracteres; después llama lethe_recall(handle) o lethe_recall("palabras") para traer el texto completo. Nada se borra nunca — el recall es sin pérdida. El skill guía le dice al agente cuándo hacerlo, así que ocurre casi-automáticamente. Es agnóstico al proveedor: funciona en cualquier host MCP (Claude Code, Codex).

2. Como librería Python (GC de contexto dentro del loop). ContextManager corre una pasada automática dentro de tu loop: un Curator heurístico puntúa cada bloque (recencia, si bloques posteriores lo citan, tipo de bloque, más una llamada opcional a un modelo barato) y un Archivist pagina los bloques más fríos a un almacén — sin pérdida, dejando stubs — para mantener el working set bajo un presupuesto de tokens. Los handles referenciados se repaginan bajo demanda.

Seamos claros con el mecanismo. Un servidor MCP no puede reescribir en silencio la ventana de contexto del host. LETHE funciona dándole al agente tools explícitas de offload/recall más un skill que hace que usarlas sea casi automático — no por magia. Esa honestidad es el punto.

✅ Qué funciona hoy vs. 🗺️ qué está en el roadmap

El diseño multi-proveedor, ensamble y retrieval semántico de abajo es la visión a largo plazo, no lo implementado. La división honesta: