Purpose: echocache is an MCP server used to cache LLM-generated results, so that conclusions already reached can be reused in subsequent sessions or between different agents instead of being re-derived. The design principle is to cache "conclusions reached by the agent," not "files the agent has read"—caching file reads cannot save tokens because the content still needs to enter the context.
Main capabilities: It provides two lookup paths. First, exact-match, HTTP-like cache_get / cache_set, keyed by (model, prompt, params), with ttl_seconds and stale_while_revalidate_seconds similar to Cache-Control's fresh/stale/expired semantics. Second, knowledge graph recall: each stored entry is a node in a small similarity graph, cache_query finds related entries semantically, cache_related traverses along graph edges (automatically associated similar entries or explicit derived-from parent nodes), and cache_invalidate can cascade invalidation along derived-from edges.
Usage: It runs as a standard stdio MCP server and can be registered directly from npm, with no need to clone or build. The startup command can be configured in Claude Desktop, Cursor, VS Code, OpenCode, or other MCP-supporting hosts; different hosts differ only in their configuration files. The repository also includes skill directories (.claude/skills/, .opencode/skills/), which hosts that support project skills will discover automatically. All settings are optional environment variables; the database directory permissions are 0700 and files 0600, and an encryption key can be set to encrypt entry contents.
Applicable scenarios: Reuse of an expensive research or judgment call, re-locating understanding across sessions, and sharing derived results across projects or hosts. The material explicitly states: it is not suitable as a replacement for file reading, is not a source of truth, and does not replace prompt caching within a single conversation.
Limitations: For a single user, a single host, and a single project, Claude Code's built-in memory can already cover most of the value for free; cross-project sharing, semantic recall, explicit freshness, and hash invalidation are its differentiators. The material states that cross-session reuse has not yet been proven to actually occur in its own history, and cross-project reuse cases remain to be measured. Enabling encryption requires a brand-new database, does not support in-place migration, and a key that does not match the databasewill be rejected at startup.
