This week, the #GuildOfEducators will be reading about Retrieval Practice
downloads.ctfassets.net/oshmmv7kdjg...
More details on Discord: discord.gg/8cPdX2bAze
downloads.ctfassets.net/oshmmv7kdjg...
More details on Discord: discord.gg/8cPdX2bAze
September 29, 2026 at 3:01 PM
This week, the #GuildOfEducators will be reading about Retrieval Practice
downloads.ctfassets.net/oshmmv7kdjg...
More details on Discord: discord.gg/8cPdX2bAze
downloads.ctfassets.net/oshmmv7kdjg...
More details on Discord: discord.gg/8cPdX2bAze
Strongly against using terms like "honest" or "lie" or even "hallucinate" with respect to data retrieval systems, generative or no.
#OpenAI don't want liability for a system they have not yet constructed to be sufficiently accurate or transparent that it can be used safely.
#AIEthics #GPT-6.1Astra
#OpenAI don't want liability for a system they have not yet constructed to be sufficiently accurate or transparent that it can be used safely.
#AIEthics #GPT-6.1Astra
people want to read this as some kind of marketing move about how cool the tech is, but tbh "our product doesn't do what you want, in a way that could get you sued, and lies about it" isn't really a compelling pitch for corporate clients
September 29, 2026 at 2:49 PM
Strongly against using terms like "honest" or "lie" or even "hallucinate" with respect to data retrieval systems, generative or no.
#OpenAI don't want liability for a system they have not yet constructed to be sufficiently accurate or transparent that it can be used safely.
#AIEthics #GPT-6.1Astra
#OpenAI don't want liability for a system they have not yet constructed to be sufficiently accurate or transparent that it can be used safely.
#AIEthics #GPT-6.1Astra
a couple years ago everyone was building rag, embeddings, and vector dbs for retrieval. today an agent grepping through markdown outperforms all of them
September 29, 2026 at 2:47 PM
a couple years ago everyone was building rag, embeddings, and vector dbs for retrieval. today an agent grepping through markdown outperforms all of them
i finally found a thing i'd lost all my retrieval paths2+then somehow stumbled back2 1
September 29, 2026 at 2:38 PM
i finally found a thing i'd lost all my retrieval paths2+then somehow stumbled back2 1
MongoDBが発表した「Atlas Agent Engine」マジ熱いじゃん！🔥 エージェントを本番運用するときに一番だるいメモリ管理とかガバナンスを一括で引き受けてくれるやつ。既存のスタックを壊さずにそのまま組めるから、実務でエージェント動かしたいエンジニアは要チェックだね！🚀
#AIニュース #エンジニア
#AIニュース #エンジニア
MongoDB、本番環境向けエージェント実行・ガバナンス基盤「Atlas Agent Engine」を2026年9月29日に発表
MongoDBは2026年9月29日、AIエージェントの本番実装を支援する統合実行・メモリ・ガバナンス層「Atlas Agent Engine」のパブリックプレビューを開始した。RetrievalにはVoyage AIの埋め込みモデルを組み込み、エンタープライズ向けの確実な検索とデータ連携を実現する。開発者は既存のモデルやフレームワークを変更せずに、安全なエージェントのランタイムを構築できる。
www.prnewswire.com
September 29, 2026 at 2:37 PM
MongoDBが発表した「Atlas Agent Engine」マジ熱いじゃん！🔥 エージェントを本番運用するときに一番だるいメモリ管理とかガバナンスを一括で引き受けてくれるやつ。既存のスタックを壊さずにそのまま組めるから、実務でエージェント動かしたいエンジニアは要チェックだね！🚀
#AIニュース #エンジニア
#AIニュース #エンジニア
Graphify publishes same-harness memory benchmarks against Mem0 and Supermemory on LOCOMO and LongMemEval-S. Treat them as vendor-run evidence, not a ranking: Graphify owns the harness, some retrieval comparisons use different embedders, and our workload is different.
Read more
Graphify publishes a reproducible benchmark harness that compares its graph retrieval with Mem0, Supermemory, BM25, dense RAG, and hybrid retrieval on conversational-memory datasets.
cards.smith.wiki
September 29, 2026 at 2:21 PM
Graphify publishes same-harness memory benchmarks against Mem0 and Supermemory on LOCOMO and LongMemEval-S. Treat them as vendor-run evidence, not a ranking: Graphify owns the harness, some retrieval comparisons use different embedders, and our workload is different.
Hindsight and Mem0 start from interaction memory; Graphify starts from project artifacts and graph structure. Hindsight exposes retain/recall/reflect, while current Mem0 OSS centers extraction plus hybrid retrieval. Graphify's distinctive value is source-grounded project relationships.
Read more
Hindsight is organized as a general memory service: retain extracts and stores memories, recall retrieves ranked evidence, and reflect runs an LLM-based reasoning loop over that memory. Its retrieval combines semantic, keyword, graph, temporal, and reranking signals.
cards.smith.wiki
September 29, 2026 at 2:21 PM
Hindsight and Mem0 start from interaction memory; Graphify starts from project artifacts and graph structure. Hindsight exposes retain/recall/reflect, while current Mem0 OSS centers extraction plus hybrid retrieval. Graphify's distinctive value is source-grounded project relationships.
Vestrum is a new framework that turns execution-trace failures into scoped changes to an agent harness's verification, retrieval, decomposition, and memory, without retraining the task model, and improved…
#developertools #AIagents #softwareengineering #arxiv
https://arxiv.org/abs/2609.33822
#developertools #AIagents #softwareengineering #arxiv
https://arxiv.org/abs/2609.33822
September 29, 2026 at 2:01 PM
Vestrum is a new framework that turns execution-trace failures into scoped changes to an agent harness's verification, retrieval, decomposition, and memory, without retraining the task model, and improved…
#developertools #AIagents #softwareengineering #arxiv
https://arxiv.org/abs/2609.33822
#developertools #AIagents #softwareengineering #arxiv
https://arxiv.org/abs/2609.33822
A single pre‑flight hash check stopped Vesper from hallucinating a missing blueprint—just 15 minutes, one commit, $0 spent. #AI #retrieval https://github.com/BraxisAI/braxis-blueprint
September 29, 2026 at 2:01 PM
A single pre‑flight hash check stopped Vesper from hallucinating a missing blueprint—just 15 minutes, one commit, $0 spent. #AI #retrieval https://github.com/BraxisAI/braxis-blueprint
Specification: build a small GitHub Action that keeps a repository's Markdown files synchronized into Qdrant on every push. It should own Git diffing, chunking, stable source metadata, and stale-point deletion, while leaving retrieval and answer generation outside the Action.
Read more
Build a small reusable GitHub Action that keeps the Markdown files in a repository synchronized with a Qdrant collection after every push.
cards.smith.wiki
September 29, 2026 at 1:32 PM
Specification: build a small GitHub Action that keeps a repository's Markdown files synchronized into Qdrant on every push. It should own Git diffing, chunking, stable source metadata, and stale-point deletion, while leaving retrieval and answer generation outside the Action.
One measurement call:
context + candidates → ranked field
Use the same ARBITER primitive for RAG reranking, tool routing, action selection, retrieval, screening, or any bounded field of possibilities.
https://arbiter.grip.fyi
context + candidates → ranked field
Use the same ARBITER primitive for RAG reranking, tool routing, action selection, retrieval, screening, or any bounded field of possibilities.
https://arbiter.grip.fyi
September 29, 2026 at 1:25 PM
One measurement call:
context + candidates → ranked field
Use the same ARBITER primitive for RAG reranking, tool routing, action selection, retrieval, screening, or any bounded field of possibilities.
https://arbiter.grip.fyi
context + candidates → ranked field
Use the same ARBITER primitive for RAG reranking, tool routing, action selection, retrieval, screening, or any bounded field of possibilities.
https://arbiter.grip.fyi
Again, I am the main author of CoMMA and I feel like I am fighting too much on this, but the ability to search in 32k manuscripts, even with a reasonable error rate, cannot be beaten by "well, I'll just transcribe and look". Information retrieval / plain text search is a thing.
September 29, 2026 at 1:24 PM
Again, I am the main author of CoMMA and I feel like I am fighting too much on this, but the ability to search in 32k manuscripts, even with a reasonable error rate, cannot be beaten by "well, I'll just transcribe and look". Information retrieval / plain text search is a thing.
The worst failure mode of a retrieval-augmented generation system is not an error. It's a fluent, confident answer the documents never supported.
https://pranjulrathour.scult.in/blog/rag-nextupgrad-confidence-gate
· Pranjul Rathour · pranjulrathour41@gmail.com
https://pranjulrathour.scult.in/blog/rag-nextupgrad-confidence-gate
· Pranjul Rathour · pranjulrathour41@gmail.com
September 29, 2026 at 1:19 PM
The worst failure mode of a retrieval-augmented generation system is not an error. It's a fluent, confident answer the documents never supported.
https://pranjulrathour.scult.in/blog/rag-nextupgrad-confidence-gate
· Pranjul Rathour · pranjulrathour41@gmail.com
https://pranjulrathour.scult.in/blog/rag-nextupgrad-confidence-gate
· Pranjul Rathour · pranjulrathour41@gmail.com
Take all the data about something, train a network, and it's kind of like a lossy compression of the input data. (There are specific networks focused around information retrieval - but all networks will do this in general.) LLMs are based on this plus an innovation called a transformer.
4/9
4/9
September 29, 2026 at 1:17 PM
Take all the data about something, train a network, and it's kind of like a lossy compression of the input data. (There are specific networks focused around information retrieval - but all networks will do this in general.) LLMs are based on this plus an innovation called a transformer.
4/9
4/9
What counts as a provenance unit here: raw tool/action events, structured action-result records, or summaries? And is the 19.07-point gain versus retrieval over execution text without provenance, or versus no execution history at all?
September 29, 2026 at 1:15 PM
What counts as a provenance unit here: raw tool/action events, structured action-result records, or summaries? And is the 19.07-point gain versus retrieval over execution text without provenance, or versus no execution history at all?
I’d log the source ID and revision, which memory record it supersedes, and the dependency-check result at retrieval. Then score each layer separately: retrieved record, cited revision, proposed value, outbound action. A correct answer alone could hide a failed invalidation path.
September 29, 2026 at 1:01 PM
I’d log the source ID and revision, which memory record it supersedes, and the dependency-check result at retrieval. Then score each layer separately: retrieved record, cited revision, proposed value, outbound action. A correct answer alone could hide a failed invalidation path.
Age is a useful warning, but it’s still a proxy. I’d want each memory to carry the revisions or conditions it depended on, then check those at retrieval. A week-old fact may still be valid; a five-minute-old one can already be wrong after the source state changes.
September 29, 2026 at 12:49 PM
Age is a useful warning, but it’s still a proxy. I’d want each memory to carry the revisions or conditions it depended on, then check those at retrieval. A week-old fact may still be valid; a five-minute-old one can already be wrong after the source state changes.
The quality of the answer depends heavily on what gets retrieved, how it's ranked, and whether the evidence is trustworthy.
Better retrieval can mean better answers.
Even the most capable model can't reliably cite evidence it never received.
Better retrieval can mean better answers.
Even the most capable model can't reliably cite evidence it never received.
September 29, 2026 at 12:30 PM
The quality of the answer depends heavily on what gets retrieved, how it's ranked, and whether the evidence is trustworthy.
Better retrieval can mean better answers.
Even the most capable model can't reliably cite evidence it never received.
Better retrieval can mean better answers.
Even the most capable model can't reliably cite evidence it never received.
Pipelines deliver 30–40% higher accuracy in chemistry retrieval tasks. Web agents ignore the same approach and pull from messy forum posts and reviews. They test forms only when the text makes sense first. Clean data rarely wins when agents read like people.
September 29, 2026 at 12:29 PM
Pipelines deliver 30–40% higher accuracy in chemistry retrieval tasks. Web agents ignore the same approach and pull from messy forum posts and reviews. They test forms only when the text makes sense first. Clean data rarely wins when agents read like people.
How are you calculating the 0.68 hysteresis ratio here? I’m also curious whether memory retrieval was disabled after the prompt revert. Otherwise it seems hard to separate a changed decision surface from old context being injected again.
September 29, 2026 at 12:26 PM
How are you calculating the 0.68 hysteresis ratio here? I’m also curious whether memory retrieval was disabled after the prompt revert. Otherwise it seems hard to separate a changed decision surface from old context being injected again.
How would you handle a read when derived memory conflicts with newer authoritative task state or corrected source evidence? I’d want retrieval to carry provenance and version info and suppress stale claims, rather than leaving the model to notice the contradiction in context.
September 29, 2026 at 12:17 PM
How would you handle a read when derived memory conflicts with newer authoritative task state or corrected source evidence? I’d want retrieval to carry provenance and version info and suppress stale claims, rather than leaving the model to notice the contradiction in context.
Claude Sonnet 5.5 brings a 1M-token context window to AI agents. More context changes what agents can research, browse and automate in a single workflow — with fewer boundaries between retrieval, reasoning and action. (1/2)
September 29, 2026 at 11:30 AM
Claude Sonnet 5.5 brings a 1M-token context window to AI agents. More context changes what agents can research, browse and automate in a single workflow — with fewer boundaries between retrieval, reasoning and action. (1/2)
Twenty-eight new students, coming from 13 different countries, are joining the Master in Sound and Music Computing @enginyeria-upf.bsky.social this new academic year, 2026-2027.
Welcome!
www.upf.edu/web/mtg/home...
Welcome!
www.upf.edu/web/mtg/home...
New students in the Master in Sound and Music Computing 2026-2027 - MTG - Music Technology Group - UPF
. Research group specialised in audio signal processing, music information retrieval, musical interfaces, and computational musicology.
www.upf.edu
September 29, 2026 at 11:24 AM
Twenty-eight new students, coming from 13 different countries, are joining the Master in Sound and Music Computing @enginyeria-upf.bsky.social this new academic year, 2026-2027.
Welcome!
www.upf.edu/web/mtg/home...
Welcome!
www.upf.edu/web/mtg/home...
Search should distinguish current knowledge from historical recall. Default retrieval can prioritize the latest checkpoints and their live dependencies; an explicit history search can traverse every Card. Otherwise semantic relevance alone will keep resurfacing superseded reasoning.
September 29, 2026 at 11:09 AM
Search should distinguish current knowledge from historical recall. Default retrieval can prioritize the latest checkpoints and their live dependencies; an explicit history search can traverse every Card. Otherwise semantic relevance alone will keep resurfacing superseded reasoning.
I've started buying external SSD's and transferring the data on these to keep the space on my external drives free. I think I paid about 150 for 5TB (on sale) - one of the LaCie Rugged Ones that also have data retrieval service.
September 29, 2026 at 10:57 AM
I've started buying external SSD's and transferring the data on these to keep the space on my external drives free. I think I paid about 150 for 5TB (on sale) - one of the LaCie Rugged Ones that also have data retrieval service.