This category actually covers two separate problems: reducing outputs that bloat the context window (compression) and retaining project knowledge across sessions (persistent memory). Token efficiency also operates at multiple layers: shell output, the model’s own responses, and tool results are targeted by different tools. Enabling two hooks in the same layer can undo the savings. Anthropic’s context engineering guide also emphasizes that the real goal is “the smallest, highest-signal set of tokens” [7].
Pay particular attention to the quality of the evidence in this area. Only two tools have undergone independent paired testing, and both results fell far short of their advertised claims: Caveman saved 8.5% [1], while RTK delivered zero or negative savings [2]. The other tools’ figures are based on their own benchmarks, so the ranking below weighs design quality, transparency, and privacy defaults. ECC from your list has been moved to category 1; the other four tools are included here.
A library, proxy, and MCP server for local, reversible compression of agent context
An automatic, persistent memory plugin that preserves context across agent sessions
MCP server and plugin that processes large tool outputs in isolation and preserves session memory
Local-first MCP memory for sharing sessions and project decisions across coding agents
A skill that shortens agent responses; a CLI that compresses tool output too
An SDK and self-hostable infrastructure for adding persistent memory to agent applications
Shared persistent memory for coding agents, backed by Markdown and hybrid search
A context engine that gives agents persistent memory, user profiles, and hybrid RAG
A skill package that teaches agents context engineering, memory, and evaluation design
A local-first MCP server for persistent agent memory in Markdown files