A no-RAG agent memory system with plain markdown saves ~155k tokens per run
gregbarbosa · x · 2026-10-07
Developer gregbarbosa built a memory system for his work agents using plain markdown files instead of RAG. It injects at most 12k tokens into the context window while saving 155k tokens compared to running without memory. He also notes agents can launch subagents with direct and referenced context to knock out multiple projects quickly—so effective he considered ripping it out but kept it. The core idea: externalize long-term context into injectable structured files rather than vector retrieval.
Related event: Developer Ditches RAG, Saves 155K Tokens with Markdown Agent Memory(2 posts)→
More from coding & agent
- Reddit thread hunting for the most private coding harness: no telemetry, no fallback models — shittihs1 · 2026-10-07
- Where Do Teams Review Agent Behavior Changes After an Eval Finds a Problem? — Afraid_Aardvark4269 · 2026-10-07
- RSI Arena Teams With Hugging Face: Agents Train Models, Community Votes on Results — my_cat_can_code · 2026-10-07
- Steal from data engineering: DAGs, lineage and incremental processing for AI agent systems — LangChain · 2026-10-07
- 40 x $90 Refunds Are Each Fine but Total $3,600: Capping Aggregate Agent Damage — No-Conflict4823 · 2026-10-07
- Matt Pocock: making agents strategic thinkers is 'really hard', often yields slop — mattpocockuk · 2026-10-07