Give Every Agent You Run in Omnigent a Persistent Memory. Most agent memory integrations solve one problem for one tool. Add Hindsight to Cursor, and Cursor remembers. Add it to Aider, and Aider remembers. Each is its own setup, its own configuration, its own ceiling. Omnigent breaks that pattern. It is a meta-harness: a single orchestration layer that wraps and coordinates multiple AI agents at once, including Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi. When you add Hindsight to Omnigent, you add persistent memory to all of them through one place, with a single setup instead of wiring up each tool's integration separately. And for any harness that has no native Hindsight integration, including custom ones, Omnigent is the memory layer. The bridge lives in Omnigent, so the agents inside it do not have to carry it themselves. Know more: lnkd.in/gJcftPPF

See the Exact Prompt Before You Spend a Token. You write a mission for your memory bank. Something like "focus on billing decisions and who approved them." You save it, and the extraction gets better, or it doesn't. Now: where did that sentence go? If you guessed the system prompt, you're wrong for two of the three operations that use it. TL;DR - A retain or consolidation mission lands in the user message, not the system prompt. Reflect's lands in the system prompt, where it replaces the agent's role line rather than adding to it. - POST /banks/{id}/prompts/preview returns the exact messages retain, consolidation or reflect would send. No LLM call, no writes. - Active blocks concatenate byte-for-byte into the real prompt, and there's a test pinning that against the actual builder. - A setting you switched off comes back as an inactive block, positioned where it would have landed. - The endpoint refuses to accept your sample text or a candidate mission, on purpose. - Shipped in v0.10.0 for self-hosted, and coming soon to Hindsight Cloud. Know more: linkedin.com/pulse/see-exact…

8 Agent Memory Use Cases (With Real Examples). "Agent memory" sounds abstract until you notice it is the same fix applied to a dozen different problems. Any time an agent should remember something past the current session, and today starts from a blank slate, that is a memory-shaped hole. Hindsight fills it with the same three operations everywhere: retain to write, recall to read, reflect to reason. What changes from one use case to the next is not the primitive, it is who the memory belongs to and where it surfaces. Here are the agent memory use cases people actually build, each with a concrete example: lnkd.in/gqe9TWR2

We Added Jev as a Reranker. Here's What We Learned. Hindsight 0.10.1 adds Jev as a reranker provider. Jev is an unusual model: it produces no text at all. You send it a passage and a typed question, and you get back a typed answer with a calibrated probability. TypeSafe calls this class of thing a System One model, after Kahneman's fast, automatic kind of thinking. That shape is a good fit for reranking, which is not a writing task. There is a pile of candidate memories and a question about which ones matter, and the only thing you want back is an ordering. Getting there took longer than we expected, and two of the designs we tried failed outright. Here is what we learned. Know more: linkedin.com/pulse/we-added-…

NVIDIA Nemotron 3.5 Embed came first on all seven retrieval benchmarks we evaluate Hindsight on. Every recall starts with the embedding model. A better one hands everything downstream a better set of memories. Try it in Hindsight, open source memory for AI agents.

Document search used to match only the document id. It now searches document text and tells you which fields matched, so the search box finds a document by what it actually says. Knowledge pages and mental models also download as PDF or Markdown.

Four new providers: → ChatGPT, to run Hindsight on a ChatGPT plan → NVIDIA NIM for embeddings and reranking → ONNX, any cross-encoder reranker in-process → Anthropic Workload Identity Federation, so you can drop the static key

A tag in observation_scopes can now be a glob. [["topic:*"]] groups each memory by whatever topic: tags it carries. That matters when the tags are added during extraction, so the client cannot know them at the time it sends the request.

Observations are rebuilt, not versioned. Edit a fact and everything built on it is deleted and rebuilt. That used to leave no trace. Each document now records what changed and which rebuild it caused, so you can work out later why an observation looks the way it does.

Hindsight 1.1.0 is out. - Per-document change history - Tag globs for observation scopes - Four new providers - Knowledge page downloads - A long round of reliability work across retain, consolidation and reflect. Everyone on 1.0.x should upgrade. 🧵

A Roster of Hermes Bots Is Useless If They Forget. Nous Research just shipped Bot Mode for Hermes. Your agent profiles become a roster of named Bots, and each one carries its own role, model, skills, and profile picture. Bots can use any model and even talk to each other in a shared room. Build a specialist once, and use it forever. The line in that announcement that matters most for us is a small one: each Bot has its own memory. That is exactly the seam where Hindsight fits, and where a roster of bots either becomes a genuine team or a set of goldfish with names. Know more: linkedin.com/pulse/roster-he…