Inject & compress 500K+ tokens without losing context

Injects large amounts of external context (documents, RAG results, prior conversations) into an agent's working memory - up to 500K-1M tokens per the skill's own claim - while deduplicating overlapping content, preserving source attribution, and aiming to prevent hallucination drift from context pollution. As a conversation grows, it automatically summarizes older turns while trying to preserve tone (sarcasm, formality, urgency), intent (what the user actually wants versus what they literally said), key facts (numbers, names, decisions, commitments), and emotional register (frustration, excitement, confusion). It compresses multi-turn history - the skill's own example compresses a 14-turn conversation into roughly 800 high-density tokens that retain full semantic meaning and can be re-expanded later - and includes a built-in fact-verification pass that cross-references controversial or ambiguous claims within the conversation and flags contradictions or unsupported assertions.

Use it when an agent is losing context in long conversations (50+ turns), when injecting large RAG or external-document context, when conversation history needs compressing without losing meaning, when fact-checking claims across a long thread, or for any agent that needs to remember everything across a session. Best practice is to enable it at the start of long-running agent sessions, turn on auto-summarization once a conversation passes about 20 turns, compress before hitting the context window limit rather than after, and run the fact verifier on high-stakes outputs; do not inject unvetted external content without deduplication, and do not skip summarization in favor of raw truncation.

Input is external context (documents, RAG results, prior conversation turns) plus the ongoing conversation history; output is a deduplicated, source-attributed context injection, a periodically re-summarized conversation history, a compressed high-density token sequence that can be re-expanded, and a fact-verification report flagging contradictions or unsupported claims. Install:

npx skills add christopherlhammer11-ai/recallmax

Distributed through the Genesis Agent Marketplace and installed via the npx skills add command; pairs with a related tool-call reliability skill, also free from the same marketplace.

Anyone running long agent sessions or injecting large external context who needs the agent to retain tone, intent, and key facts across dozens of turns instead of losing them to truncation or drift, and to have contradictory or unsupported claims flagged automatically.