RAW-RAG: keep a raw-history baseline when evaluating agent memory

Preserve evidence before interpreting it

The implementation converts overlapping passages into searchable numerical representations, retrieves candidate evidence, reorders it by relevance and asks a reader model to answer. Timestamps and relations can matter as much as isolated facts. The authors report leading six of seven compared settings, with a loss in one MEME conversation-memory evaluation.

Is Complex Agentic Memory Always Needed? ↗RAW-RAG implementation and results ↗

Compare like with like

Repository timing excludes answer generation and evaluation. The public paper discussion asks about retrieval costs; the author describes both hosted and open-model alternatives. Those comments clarify the setup but do not independently reproduce it.

RAW-RAG implementation and results ↗RAW-RAG discussion and author response ↗

A small acceptance example

Create a fictional history: delivery is Tuesday, changes to Thursday, then is cancelled. Ask for the final status and the passage that changes it. Compare raw retrieval and summarized memory on the same questions, reader and budget. Include cases where an earlier fact becomes invalid.

RAW-RAG implementation and results ↗

What you can do

1. Inspect the benchmark table and timing exclusions before comparing this baseline with an existing memory service. Open reference ↗

Where the evidence stops

The benchmark results do not prove raw retrieval wins for every workload or smaller reader model. The fixture above is original and unexecuted.

Sources read. Checked October 23, 2026. The paper, repository and author discussion were read. No memory benchmark or original fixture was executed.

Sources & further reading