Notes on company memory.
Research and practical thinking about decisions, knowledge loss and the work it takes to recover what a company already knew.
Companies Do Not Have a Search Problem
A company does not lose knowledge because it cannot find files. It loses knowledge because it never kept the reasoning. The anchor piece: what a company memory has to be, and the four beliefs it has to hold.
Company Brain Ownership: Rent the Model, Own the Record
Company brain ownership is the new contract test: rent the model, but export the record, its corrections and its permission map, or you do not own it.
Agent Memory Portability: Why the Next Fight Is About Moving Memory, Not Storing It
Agent memory portability is the new test. Graphon raised $35M, Redis shipped GA memory, and LoreKit says the formats cannot read each other.
Teamwork Graph: Atlassian Maps 250 Billion Objects. Who Keeps the Graph True?
Atlassian's Teamwork Graph maps 250 billion objects and 10 million monthly human-agent collaborations. A bigger graph is not yet a true record.
Agentic Ontology: Databricks Put the Company Map in Charge of an Always-On Agent
Databricks made its agentic ontology generally available and launched always-on Genie. The tool for curating that map is still heading to beta.
Prebuilt AI Agents: IBM and Microsoft Offer 6,000. Adaptation Is the Work.
IBM and Microsoft launched Imagine Enterprise on Nov. 10 with 6,000+ prebuilt AI agents. The catalogue is instant. The company record is not.
Memory Trap: Benign Notes Raised Unsafe Agent Actions From Under 1 Percent to 81 Percent
The memory trap: benign agent notes raised unsafe actions to 81 percent in a Nov 5 study. A memory needs its scope, not just its text.
Knowledge Compilation: Anthropic, Google and RAGFlow Compile the Record Before the Question
Knowledge compilation moved enterprise context ahead of the question. Anthropic, Google and RAGFlow show what a compiled record must prove.
Fragmented Enterprise Knowledge: Why Only 10 Percent of Companies Let AI Agents Act Alone
TCS surveyed 1,333 leaders: 84% cite AI productivity, 10% let agents act alone, and 44% say knowledge is fragmented. Autonomy follows the record.
Context Boundary: Claude Code Sent the User Email to Websites. The Leak Was in the Context.
Claude Code put account emails in User-Agent headers. The fix filters one shape. A context boundary must hold outside the model, on every tool call.
Context Retrieval: Box Promises 90 Percent Fewer Tokens. The Passage Is Not the Record.
Box Context Retrieval claims 20 percent better recall and 90 percent fewer tokens. Retrieval saves cost, but a passage still needs a record behind it.
Official Authority Layer: RC AI Brain Marks Which Company Record Wins
RC365's RC AI Brain beta adds an official authority layer tagging version, effective date and scope. Marking a winner is not keeping it true.
Agent Memory Correction: Salesforce Doubled Recall by Replacing Facts, Not Adding Them
Salesforce doubled Agentforce memory recall by treating updates as corrections. Why correcting facts, not storing more of them, decides if a record stays true.
Persistent Agent Memory: Perplexity Brain Shows Recall Is Not the Test
Perplexity's Brain guide shows agent memory as a wiki that rewrites itself. The gains are real, the test scores are mixed, and the real test is a fresh start.
Mosaic Leak: Every Document Was Permitted. The Answer Was Not.
USC tests found agents disclose conclusions assembled from permitted documents. A mosaic leak passes every access check. The defence is to test the combined answer, not each document on its own.
Memory Tax: Raw Search Beat Agent Memory Systems in 90 of 99 Tests
Two October studies found plain search beat agent memory in 90 of 99 tests: facts get lost when memory is pulled out of documents. That loss is the memory tax.
Agent Delegation Failures: 23 Ways Agents Fail the Person Who Asked
A study of 163,487 Reddit posts lists 23 ways agents fail the person who asked. The worst are the hardest to spot. The fix: keep a record of what was handed over.
Why AI Agents Need to Know Who Can Do What, Not Just Documents
Docebo tied its knowledge layer to skills data on Oct 28. Only 34% of agent projects ever get into real use. Why trust depends on the record behind the layer.
Operating Knowledge: When Agents Act on What a Company Knows, Not Just Answer
Fragment raised $56M so agents can act on what a company knows, not just answer from it. When agents act, the record they act from becomes a risk in itself.
AI Tools That Do Not Share Stall Work for 59% of Executives
Salesforce finds work stalls weekly for 59% of executives because their AI tools do not share what they know. Shared knowledge only pays off as a record people have reviewed.
Changing Records Break AI Agents: Top Model Scores 32.5% When Facts Change Over Time
A new test, MMEnterprise-Bench, scores the top model at 83.3% on a single document and 32.5% when records change over time. Better search alone does not fix it.
Agent Memory Benchmarks: Same Answers, Scores 18 Points Apart
Agent memory scores swing 12 to 18 points when the grader changes, with the same answers. What to ask before trusting a memory score.
Zoom Buys Onyx, Neo4j Builds the Map: One Record for Every Agent
Zoom bought Onyx on Oct 23 and Neo4j launched its Context Engine the same week. Both bets say the same thing: agents need one well-kept record of company knowledge, not more search.
Shared Agent Memory Is Here. The Record It Creates Is the Hard Part
Databricks and CollectivIQ launched shared agent memory in the same week. The gains are real, but a shared record only works if every fact carries its source and a person reviews it.
Models That Edit Their Own Memory, and the Hard Part They Leave Behind
Models are starting to manage their own memory, with a dedicated memory writer inside. The gains are real. A company's record is a different problem.
26,000 Agents at Volkswagen: Enterprise AI Agent Deployment Hits a Memory Problem
Volkswagen gave 330,000 staff Copilot and ran 26,000 agents in a month. At that scale, the hard part is memory: what agents remember, and whether it is right.
Google Makes AI Agents Declare Themselves by November 11
Google will block Workspace MCP calls from apps not marked as agent apps on November 11. A label says which traffic comes from agents. It does not control what they do.
The AI Verification Burden: Agents Write Faster. Humans Still Have to Check.
AI agents produce 30% more code but no more finished work, Harvard data shows. Whether agents pay off now depends on how much checking their work takes.
Agent Identity: Microsoft Gave Agents Their Own Accounts. The Memory Is Shared.
Microsoft's Oct. 15 agent accounts give agents their own logins and a shared memory. The data shows why every agent still needs a named person who owns it.
Microsoft's Team Knowledge Base Makes Agents Faster. Who Keeps It True?
Microsoft's new team knowledge base makes agents up to 4x faster in its own tests. The gain comes from doing the gathering once. The risk is who approves what goes in.
Shadow AI Agents: 1,000 Tools Inside the Enterprise, Outside Its Controls
About 1,000 outside AI agents run inside company software, outside the company's sign-in controls. The risk is not only security. It is what they read and keep.
Memory Poisoning: The Attack That Turns AI Agent Memory Into a False Record
Memory poisoning plants false facts in an agent's memory, and they stay there. The research, the numbers, and the defence: every fact carrying its source.
Why AI Agents Need a Map of Decisions, Not Just Documents
Only 34% of agent projects ever get into real use. Context graphs record why decisions were made, not just what files say. Here is the evidence.
What a Company Knows: AI Agents Need It Written Down, Not Just Memory
A Hacker News debate argues agents need written documentation, not memory add-ons. The lesson: records a person has reviewed come first.
How Jilu Works
How Jilu turns chat, documents, tickets, meetings and pages kept from the web into one company record: what gets collected, who can see it, how decisions are pulled out and reviewed, and answers that show their sources.
The AI Agent Memory Race Is On. Companies Are Behind.
Google's Gemini agent ships with four kinds of memory. New Templafy data shows agents that do not know the company cost workers hours of redoing work.
Enterprise Search Is Getting More Expensive: Atlassian Raises Prices Up to 10%
Atlassian Cloud list prices rise up to 10% on October 13, 2026. Why enterprise search pricing is under the microscope, and what the market data says.
The Company Brain Gold Rush: Why AI Agent Memory Is Drawing Hundreds of Millions
Engram's $98 million launch caps a funding wave built on one idea: agents that cannot remember cannot work.
Search finds files. Memory finds the reason.
The difference between retrieving a document and reconstructing the decision it changed.
Why postmortems work, and what every team can learn from them.
Google's incident practice shows how a shared record turns one failure into company-wide learning.
When someone leaves, what exactly walks out?
Turnover does not only remove capacity. It changes who can reach the knowledge behind the work.