MICROSOFT QUIETLY BUILT ONE MEMORY MODULE THAT PLUGS INTO ANY AI AGENT AND CUTS ITS CONTEXT BY UP TO 100X no retraining. no task-specific redesign. you just plug it in Microsoft's insight: agents don't need to remember what happened. they need the knowledge pulled from it - the facts... and the reusable skills, not the raw logs so PlugMem stops storing history. it compiles experience into a knowledge graph, then feeds the agent only what matters the result: > up to 1-2 orders of magnitude fewer tokens > highest decision-utility per token of any system tested > one unchanged module beats task-specific memory on chat QA, multi-hop search, and web agents more memory was never the answer. better memory is - and Microsoft just shipped it as a plug-in bookmark this. the agents that win won't remember more, they'll remember smarter.show more

per grok, the repo for this is located --- The public code repository for PlugMem is here: It is maintained by the TIMAN group at the University of Illinois Urbana-Champaign (with co-authors from Microsoft Research and Tsinghua). Both the Microsoft Research blog and the arXiv paper (2603.03296) link to this same GitHub repo for the code, experimental results, and plugins. There does not appear to be a separate official Microsoft GitHub organization repository for PlugMem—the public implementation lives under TIMAN-group.

Seems a hell of a lot like Obsidian

The next generation of AI memory may be less about storing everything and more about extracting what matters.

compression schemes for agent memory all lose the same thing first, the attempts. you keep the conclusion and drop what was already tried, so it quietly retries the path that already failed a week later. i tried Runner, it logs the failed attempts too, written with ai

Fuck I was building this

The most misleading sentence is “PlugMem stops storing history.” It does not. It separates history from decision-ready knowledge and normally sends only the distilled knowledge into the agent’s active context. The second major omission is memory-processing cost. PlugMem can substantially reduce inference-time context while incurring considerable upfront LLM work to construct and maintain the graph. Whether it reduces total cost depends on how often that memory is reused.

Agree. Agents don’t need to remember more; they need to know how to forget. RAG is passive. I built Hypotree, an MCP server that wires memory as a DAG, prunes dead branches, and draws its own conclusions by process of elimination. Deep dive here:

@N01ennn The question is how the knowledge graph handles contradictions between experiences. If two past episodes yield conflicting skills, the compilation step has to pick a winner. Curious how that degrades in long-horizon agents.

Context windows solve capacity. Memory systems solve continuity. The next generation of agents will need both.

That distinction matters more than people give it credit for. Most memory hacks just summarize the conversation. The useful version separates "what happened" from "what's still true", so the agent isn't re-deriving state from a wall of past turns every time.

smarter memory >>> more memory and Microsoft solve this problem with a simple plugin

@grok 这个纯研究还是有工业意义，具体工业场景视角看意义是什么，有开源数据集或者开源项目代码吗？从多个数据源交叉验证，不要只看新闻媒体一面之辞。帮我排除没意义的垃圾商业营销推广、诈骗 以及自吹自擂，自嗨，无病呻吟。

This feels like a more scalable approach. Compressing experience into reusable knowledge makes a lot more sense than replaying entire conversations.

Compiling experience into a graph instead of hoarding logs is the right instinct. Raw history is mostly filler and replaying it is what makes long sessions worse, not better. The open question is what happens when a distilled fact turns out to be wrong, and whether anything downgrades it later.

i’d rather my agents keep the lesson than every old log

Reducing context size while improving decision quality is a much more sustainable approach than simple increasing context windows. Excited to see how ideas like knowledge graphs shape the next generation of AI agents.

Where git?

Nobody asked what happens when that memory stores something it shouldn’t. @domAIynLabs

Shifting from brute-force context windows to distilled semantic synthesis is the real breakthrough here. By extracting rules and insights instead of hoarding raw chat logs, we solve the biggest compute bottleneck in long-running agentic workflows. This is how we scale production agents.

The key design win is separating episodic logs from reusable knowledge. Agents can retain raw traces for auditing while using a compact graph for inference and action.

Will the memory somehow be editable?

This feels like the next phase of AI efficiency: not bigger context windows, but better information compression.

The future of AI memory is not storing everything. It’s about extracting the right knowledge, compressing experience, and delivering the context an agent actually needs.

What's the name of this and is it open source?

This is the direction AI memory should move toward. The goal isn’t storing everything, it’s extracting useful knowledge and delivering the right context at the right time.

Clever shift from logs to usable knowledge.

plug-in memory modules that prune static context before token injection cut inference overhead dramatically in multi-agent swarms. we benchmarked external vector caching vs embedded state pruning; dropping raw message history in favor of distilled state vectors saved ~65% latency on deep agent loops. are you testing external memory modules with dynamic graph retrieval or flat semantic search?

That's the next level in the evolution of the LLM-wiki

It would be nice as well if its editable by us users, something like CRUD

Does this process introduce bias?

The useful half isn’t “Microsoft memory plug-in.” It’s: agents don’t need the transcript — they need distilled facts and reusable skills, then only the slice that changes the next decision. More context was never the moat. Better memory on metal you own is.

reusable skills beat raw logs for agent memory

Neden değişmemiş? Madem yapıyor bu bilgiyi bulutta tutsun her bağlantıda kullansın?

my agent remembers every dumb thing ive ever asked it and none of the useful information this would genuinely make my life easier