Galahad

Galahad is an AI memory layer that saves and restores model readings to disk so GPUs never reprocess the same text.

Galahad is a memory layer for AI inference that saves a model's KV cache to disk and restores it to prevent recomputing previously read tokens. It features a C++ core and integrates directly with runtimes like llama.cpp, vLLM, and SGLang. Beyond caching, it stores exact text documents and identifies the specific chapters needed to answer questions, keeping GPU memory flat while reading texts up to 50 million tokens. It is built for developers running local or hosted LLMs and supports frameworks such as LangGraph, LangChain, and CrewAI.

Galahad at a glance

Best for: Software engineer
Runs on: Linux, API
Pricing: Free for 1 GPU.
Released: October 2026
Category: AI coding tool
Website: github.com/corbenicai/galahad

Key features

KV cache disk persistence and retrieval with Taliesin
Exact document storage and chapter retrieval with Blaise
Moving context window supporting texts up to 50M tokens
Direct C API embedding without a server or Python
Agent connectors and causal graph tracking
Agent replay via a pytest plugin for CI builds
Copy-on-write KV cache branching and snapshots
AES-256-GCM encryption at rest with local key control

Use it for

Preventing redundant token computation across repeated LLM inference requests
Reading documents up to 50 million tokens with flat GPU memory usage
Testing and debugging AI agent regressions in CI pipelines using replay
Sharing system prompt prefixes across multiple agent requests
Branching agent execution paths using copy-on-write memory states

Alternatives to Galahad

Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
AI Memory — Shared memory for coding agents that works across machines and team members.
LMDeploy — Accelerates inference with high-performance engines and quantization tools.
Turnbook — Saves every chat turn locally with zero model tokens and shares memory across AI apps.
Alook — Connects existing local coding agents to team communication channels.
fal — Access 1,000+ generative models and high-performance serverless GPUs.
Memorax Code — Carry project architecture and verified engineering decisions across coding sessions.
llm-d — Deploys production-grade LLM serving on existing hardware clusters.
Engram — Single Go binary with zero dependencies and built-in full-text search.
Cognee — Persistent memory for AI agents using graph, vector, and relational retrieval.
Soup — Fine-tune and align models like Llama-3.1-8B on a 4 GB laptop GPU.

All Galahad alternatives