Recall speeds up RAG with episodic memory, cutting hits to 260ms from 49s
sharpeye_wnl · x · 2026-07-29
A RAG memory system called Recall stores episodic memory before hitting the document store. If a previous answer is fresh and confident, it skips retrieval entirely, cutting the response path from 49s cold / 6s warm down to 260ms on memory hits.
What it does
- Uses cache + ChromaDB to detect semantically similar questions.
- Runs a memory hit path that returns a cached answer directly.
- Falls back to the normal retrieval + generation pipeline when needed.
- Also supports a reconciled path where both routes run and an LLM arbiter picks the better answer.
Claimed latency
- Memory hit: 260ms
- Retrieval + generation: 49s cold, 6s warm
- Reconciled: 5–8s
More from coding & agent
- agensis bets on shared workspaces where humans and agents collaborate in channels — jasonkneen · 2026-07-29
- Reddit discusses how agents should make first contact with humans — Seeqit-Official · 2026-07-29
- Reddit says switching inference providers cut AI agent latency sharply — Cookie_cutie_69 · 2026-07-29
- Reddit debates how to verify long-running agents without redoing their work — Seeqit-Official · 2026-07-29
- OpenWork positions itself as an open-source alternative to Claude Cowork — different-ai · 2026-07-29
- RepoSkein exposes a deterministic repo graph to coding agents over MCP — JeremyCMorgan · 2026-07-29