Was just sitting here working away and saw a text from Atom, looks like I just sold AgentMemory(.)io for $7,500. I don't sell a lot of domains these days, so always a pleasant surprise when it happens. Congrats to the new owner, excited to see what they do with it!

🚀 Hy Memory 2.0 is out — open-source model, framework, and MemHub. Your private memory steward. Framework handles chat, work & code memories. #1# open-source on 4 benchmarks. Hy-Memory-30B-A3B beats Opus 5 across major frameworks. MemHub (CLI + Mac) syncs across 10+ AI tools. 💻➡️ 🔗 SDK & MemHub: 🔗 Model: #HyMemory# #AgentMemory#

Memory Genesis Competition 2026 is in last call — submissions close on March 15. You're also welcome to join us on April 4 at the Computer History Museum for an in-person gathering and high-signal conversations with the EverMind core team and leaders across OpenAI, AWS, research institutes, open-source communities, and the investment world. Guess who will you meet? Follow the competition website for the latest updates: #AIMemory# #AgentMemory# #EverMemOS# #AgenticAI# #Hackathon# #Developers# #AIInfra#

How should you design an agent's memory? Anthropic compares it to how humans consolidate memories during sleep, and proposes "dreaming" 🧠 The components of agent memory ❓ What layers does memory have? 💡 From the bottom: transcripts of everything, memory notes written while working, structured representations such as wikis, and a small injected context loaded into every session. A wrong note in the injected layer is carried into every session, so it should stay as small as an index. ❓ What is dreaming? 💡 An offline consolidation process between sessions that reads the lower layers and updates the upper ones. It checks notes against transcripts to fix errors, merges duplicates, and promotes what matters. In Pokémon Red's Mt. Moon, the agent reached Cerulean City after a wrong note was corrected. ❓ Does it actually work? 💡 ・On CL-Bench, Opus 5 rose from 0.564 to 0.828 ・On about 350,000 Stainless files, wiki-only accuracy climbed from 13% after 2 dreaming sessions to 48% after 128 ・Combining wiki and search saved about 20,000 tokens per question Most results come from single runs, so treat effect sizes as indicative. ❓ What should the agent write down? 💡 What it can't recover from the environment: strategies and tested rules, not what the screen or tools show every turn. Fixed rules like "record only failures" shrank the gain for stronger models. #AIAgents# #AgentMemory#

For agent memory, building structure that lets you trace back to evidence may beat just adding more retrieval 🧠 Building Evidence-Grounded Agent Memory with OpenWiki and Graphs This LangChain post combines a compiled wiki (OpenWiki) with an evidence graph built from its citations to improve accuracy on questions that span multiple documents. Three highlights: 📚 A wiki tied to sources by hash Each wiki statement is linked to supporting source lines and a hash of them. When the source changes, affected claims are flagged without a model call, and only those are re-checked by a model, so the memory can self-correct. 🔗 Typed edges from existing citations Edges such as cite, supersede, and refers_to are built from existing citations and in-document references with no extra model calls. The agent just uses locate to find a starting point and expand to follow edges. The graph built in under a second. 📊 About 5x accuracy on connected questions On a synthetic corpus of 111 insurance documents, accuracy on connected questions rose from 0.10 with search to 0.52 with the wiki. Adding the evidence graph improved accuracy by 27% and cut cost by 58%. For simple lookups, search did just as well. Memory pays off when agents repeatedly reason across multiple sources. #AIAgents# #AgentMemory#

An agent's memory is really an information retrieval problem. That's the argument of this perspective paper by Qingyao Ai of Tsinghua University. Title: Agentic Memory Management Is the Next Frontier of Information Retrieval URL: 🧠 Overview IR grew up assuming a human reads the results. Agents instead use retrieved information as their own internal state. The paper organizes this as agentic memory management (AMM). ❓ Problem Agent-memory research is ad hoc, borrows inconsistent terms from human psychology, and relies on weak evaluation. 🧩 Proposed framework ・constants: facts (matched by content) ・functions: skills and workflows (matched by applicability) ・operators: preferences and style (matched by scope) Five operations (add, update, delete, search, use) are grouped into learning, forgetting and utilization. Systems such as RAG, MemGPT and Mem0 are classified, with reusable skills and principled forgetting found to be underexplored. 🔭 Research agenda ・Benchmarks covering all three layers and forgetting ・The "observability inversion": diagnosing with downstream task outcomes ・Human control and security, such as memory poisoning There are no experiments; it offers a framework and conjectures. #AgentMemory# #InformationRetrieval#

Agent memory gets good not by writing more application code, but by tuning configuration. A practical guide from Weaviate. Title: Agent Memory with Engram: A Practical Guide URL: It walks through moving from "just add and search conversation data" to actually optimizing both memory quality and token cost. Three things stand out. 📝 Write topic descriptions as exclusions The description you write for each memory category acts directly as the extraction prompt. Rather than enumerating examples of what to keep, a single exclusion rule like "do not record events, incidents, or passing conditions" generalized better to cases nobody anticipated. The description also decides the shape of the output: atomic facts versus flowing prose. 🔒 Pin singular facts with bounded Marking a topic bounded caps it at one memory per scope, so facts get updated instead of accumulating. Across five fresh users, a bounded UserProfile held at exactly one memory on all five runs, while the unbounded UserKnowledge fluctuated between two and four each time. 💰 Where you put memories decides your bill Search results change every turn, so pasting them into the system prompt breaks the cache and bills the whole prompt again. Instead, fetch always-on memories once at session start and place them right after the system prompt, then append search results after the user message. Over 25-turn sessions, the final request was about 3,500 tokens with only roughly 100 of them not served from cache. Worth reading as a decision aid if you're weighing building a memory layer yourself versus outsourcing it. #AIAgents# #AgentMemory#

🧠 Maybe an AI agent's memory shouldn't be organized the moment it's stored, but the moment it's recalled. Title: Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents (JitMem) URL: ❓ What's the core idea behind JitMem? Traditional agent memory summarizes an experience at write time, right after a task finishes. JitMem skips that: it stores raw trajectories as-is and only synthesizes a task-specific summary at read time, once the new task is actually known. ❓ Why isn't write-time summarization good enough? You have to decide what matters before you know which future task will need it, so valuable details often get discarded for good. Worse, the same experience can teach different lessons depending on the task, but a write-time summary can only be one generic version. ❓ How is it trained? A curator model that writes summaries from the current task and past trajectories is trained directly on the success reward the executor gets from using that summary. No need to wait for future queries, which makes optimization much simpler. ❓ How much does it help? Across ALFWorld, WebShop, and τ²-bench, JitMem beats the strongest baseline by 16.2, 16.3, and 3.9 points respectively. Even the untrained version is already strong, showing that read-time curation itself is a major source of the gain. #AIAgents# #AgentMemory#

Agent "memory" that only learns from past trajectories can quietly bake in wrong knowledge. Here's a smart fix for that. Title: Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents URL: ❓ What was wrong with traditional agent memory? 💡 Curating memory purely from completed trajectories means over-generalizing from a single lucky observation, and memories going stale the moment something like a database schema changes. ❓ How does this paper fix it? 💡 Before committing anything to memory, it lets the agent probe the live environment with read-only tools to verify the knowledge is actually correct, using a "propose-probe-commit" pipeline. ❓ How much does it actually help? 💡 On a database exploration benchmark, pass rate jumped from 39% to 73%, reward roughly 2.6x higher, while tool calls dropped 47% and cost fell about 50%. ❓ Is it practical for real deployments? 💡 Yes — it needs no model retraining and doesn't change the task-time interface, so it can be dropped into long-running enterprise agents as-is. #AIAgents# #AgentMemory#