A new context engine plugin, warm_compaction, speeds up Hermes Agent compaction by having the main model write the handoff summary on the cached prefix of its last request, so the server reads only new rows. It works on unpatched Hermes Agent installs using documented plugin APIs. This reduces compaction latency for agent workflows without modifying host code.
A plugin for the OpenCode coding agent that performs context compaction in a way that preserves the local model's prompt cache. Instead of stripping tokens from the conversation start, it keeps system prompt and tools in place and asks the LLM to write a summary, then replaces the rest of the conversation with that summary. This avoids costly re-prefilling of already cached conversation tokens when running local models.
A plugin for the OpenCode coding agent that performs context compaction in a way that preserves the local model's prompt cache. Instead of stripping tokens from the conversation start, it keeps system prompt and tools in place and asks the LLM to write a summary, then replaces the rest of the conversation with that summary. This avoids costly re-prefilling of already cached conversation tokens when running local models.
A lossless context compression tool for coding agents that preserves tool output (file contents, JSON, stack traces, diffs) exactly. It sits between the agent and the LLM, rewriting only blocks that can be proven exactly reconstructible, and attaches a BLAKE3 certificate for verification. Solves the problem of token waste in agent loops without losing critical information.
A lossless context compression tool for coding agents that preserves tool output (file contents, JSON, stack traces, diffs) exactly. It sits between the agent and the LLM, rewriting only blocks that can be proven exactly reconstructible, and attaches a BLAKE3 certificate for verification. Solves the problem of token waste in agent loops without losing critical information.
A new arXiv paper introduces ReCAP, a memory compaction method for LLM agents that stores attention-derived importance scores and dependency links in a persistent context graph. It selects relevant history for each new request without extra model calls, cutting compaction latency by ~95% versus Codex's summarization-based compaction and roughly halving historical context per call.
A new arXiv paper introduces ReCAP, a memory compaction method for LLM agents that stores attention-derived importance scores and dependency links in a persistent context graph. It selects relevant history for each new request without extra model calls, cutting compaction latency by ~95% versus Codex's summarization-based compaction and roughly halving historical context per call.
A wrapper for Claude Code that automatically compacts idle sessions while the prompt cache is still warm, reducing cold-start latency and usage costs. It solves the problem of wasted context and slower responses when returning to a session after a short break.
A wrapper for Claude Code that automatically compacts idle sessions while the prompt cache is still warm, reducing cold-start latency and usage costs. It solves the problem of wasted context and slower responses when returning to a session after a short break.
Supercompress is a query-aware context compression tool for LLMs and coding agents that reduces input tokens by roughly 65% while preserving evidence. It targets the cost of irrelevant context in RAG chunks, chat history, tool dumps, and logs.
Supercompress is a query-aware context compression tool for LLMs and coding agents that reduces input tokens by roughly 65% while preserving evidence. It targets the cost of irrelevant context in RAG chunks, chat history, tool dumps, and logs.