Theme 1 – Memory is just plain text/markdown with search (RAG or grep)
Many commenters see agent memory as nothing more than storing notes in markdown files and retrieving them with simple text search or semantic embeddings.
“It's markdown.” – JustFinishedBSG
“Just use text.” – thm
“It's markdown + RAG/semantic search.” – wasabi991011
“Agents … grep the whole repo for plausible keywords.” – dist-epoch
“Markdown + a simple embedding.” – calpaterson
Theme 2 – Memory quality, poisoning, and the need for curation
Several participants warn that uncurated memory quickly becomes noisy, outdated, or misleading, and must be actively maintained.
“Memory is prone to poisoning.” – gritzko
“You have to constantly tend the garden and weed these things out.” – iamflimflam1
“Nothing beats curated data. Memory should be regularly reviewed, compacted, and cleaned up.” – docheinestages
“If there is one poisoned line of text it negatively affects everything else downstream.” – dataviz1000
Theme 3 – Skepticism about the novelty/value of specialized agent‑memory specs
A recurring view is that existing tools (SQL, RAG, grep, plain prompts) already do the job, and new specifications feel like reinventing the wheel.
“Agents can already call SQL / RAG and grep through files or whatever. Why is ‘agentic memory’ a special thing?” – dominotw
“They just want to reinvent information retrieval from first principles… reinventing hexagonal wheels.” – TudorAndrei
“The tool is already bash and the optimized API is grep… we’ve been here many months.” – phoghed
“I don’t think I feel a need for more agent memory… do I need more discrete memory my team has to maintain?” – ltsSmitty