外部記憶があると Claude Code の /compact はほぼ要らないThis article explains how integrating external memory via MCP allows Claude…

Claude Code, Anthropic's terminal-based coding agent, operates within a finite context window, and long-running development sessions inevitably approach that limit. The usual remedy is the built-in /compact command, which condenses the accumulated conversation into a shorter summary to reclaim space. A recent Zenn post argues that once you connect external memory through the Model Context Protocol (MCP), that command becomes largely redundant, because the meaningful state of a session can live outside the model's working context instead of competing for room inside it.

This matters because context management is one of the more tedious aspects of working with agentic coding tools. Every token spent on old conversation history is a token unavailable for reasoning about the current task. When the window fills, the agent's responses can degrade, and /compact, while useful, is inherently lossy: summarization discards detail, and the developer must decide when to trigger it. Frequent compaction can quietly drop the specific decisions, file paths, or edge cases that mattered earlier in the work.

MCP is an open standard Anthropic introduced in late 2024 to connect language-model applications with external tools and data sources. An MCP server exposes resources and callable tools that the model can invoke during a session, ranging from file systems and databases to web APIs. Among the reference implementations is a memory server that maintains a persistent store, often structured as a knowledge graph of entities and relations, that the agent can write to and read from across sessions. The article's central idea is to treat this kind of store as the canonical record of long-term work state.

In practice, the workflow shifts from keeping everything in the conversation to deliberately externalizing it. Task lists, architectural decisions, the current state of a refactor, and notes about what has been tried are written to the memory server as the work proceeds. When the agent needs that information again, it queries the store rather than relying on it still being present in the visible transcript. Because the essential details are retrievable on demand, the context window can stay comparatively lean, and the pressure that would normally force a /compact is reduced.

The approach is not entirely automatic, and the post appears to acknowledge the tradeoffs. The model has to be prompted, through instructions or conventions, to save and retrieve at the right moments, and each retrieval consumes some tokens and adds latency. Poorly organized memory can also become noise. In that sense, external memory trades one kind of overhead, manual compaction, for another: designing and maintaining a disciplined read-and-write habit. Whether that is a net gain likely depends on the length and complexity of the projects involved.

It helps to place this alongside adjacent mechanisms. Claude Code already supports a CLAUDE.md file, a project-level document loaded at the start of a session that carries conventions and context automatically. That is effectively static, curated memory, whereas an MCP memory server is dynamic and updated as work happens; the two can complement each other. The broader pattern also echoes retrieval-augmented generation, where relevant information is fetched from an external store, frequently a vector database, and injected into the prompt only when needed, rather than being held in context permanently.

The technique fits a wider industry trend toward externalizing state for AI agents. Tools such as Cursor and other coding assistants have introduced their own memory and rules features, and a growing ecosystem of MCP servers has emerged since the protocol was published, with support now appearing in several clients beyond Anthropic's own products. Even as context windows grow larger, with Claude models offering windows in the hundreds of thousands of tokens and longer variants in testing, larger capacity does not eliminate the problem, because agentic coding can generate enormous volumes of intermediate output. Off-loading durable state to a dedicated system remains a plausible way to keep sessions