Knowledge compilation became the enterprise default in 48 hours. On 5 and 6 November 2026, Google, Anthropic, Zep and InfiniFlow each moved work that used to happen at query time to a compile step that runs before the question arrives. Instead of searching raw files on every request, their systems now build wikis, context sets, evidence chains and skills in advance, and serve the compiled result to the agent. The shift saves tokens. It also creates a new record, and that record is the part that needs governing.
Knowledge compilation in three releases
Anthropic published The components of agent memory on 6 November. The post separates four layers: searchable transcripts, notes the agent writes during work, representations such as a linked wiki, and a small layer of injected context loaded into every session. A background pass Anthropic calls dreaming checks notes against transcripts, merges duplicates and corrects errors. The results favour judgement over volume. In a Pokemon Red comparison at about 10,000 steps, one run each, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files and earned four badges. On about 350,000 internal files, about 420 million tokens, a wiki built over 128 dreaming sessions rose from 13 percent to 48 percent of graded criteria on 25 questions answered from the wiki alone. Wiki plus search reached 61 percent, against 53 percent for search alone, and used about 20,000 fewer tokens per question. On a contract task, 41 of 57 facts survived in a shared page. With one page per contract and word for word quotes, all 57 survived. Anthropic presents these as its own evaluations, and the Pokemon comparison is one run per model, not an independent benchmark. The pattern is reported in the Week 45 roundup at Semantic Tech Insights.
Google made the same move for enterprise data. On 5 November, Google Cloud announced self-learning enrichment in preview for Knowledge Catalog, which it calls a universal context engine in its Agentic Data Cloud post. The system mines Google Workspace, GitHub, Jira, Confluence and SharePoint, fills context gaps based on the questions people ask, and packages the result as reusable context sets in Open Knowledge Format. Data stewards can review, audit and approve proposed updates in a new studio experience, with full versioning and rollback. Business definitions come from trusted semantic models, now generally available, rather than from the model. Google says its internal benchmarks made agents nearly four times more accurate, cut latency by up to 33 percent and used up to 75 percent fewer tokens. Google gave no test conditions for those figures, so they are a vendor claim, not a measured result a buyer can repeat.
InfiniFlow gave the pattern its name. RAGFlow 1.0, released on 6 November, and its Context as Infrastructure post on 5 November describe knowledge compilation as moving understanding from query time to indexing time. The central artifact is an atomic proposition: a one sentence fact paired with a quote from the source, and the quote must be a contiguous piece of the original text. The team says compiled evidence can save over 90 percent of tokens on a question when it hits, and costs about 2 percent when it misses. A separate review of the RAGFlow preprint reads the evidence more narrowly: under a strong search harness, structural compilation did not raise accuracy, while compiling evidence alone cut average tokens by 38 percent. That preprint is vendor authored, based on a single run with no confidence intervals, and its authors state those limits.
Zep applied the same discipline to procedure. Agent Skills, launched on 5 November, compiles agent runs into versioned SKILL.md files. Only runs whose success was confirmed by a tool or an evaluation can support a Skill. The agent saying it succeeded does not count. Every step cites the run it came from, and deleting a source run deletes every Skill that cites it. In Zep testing, skills compiled from Claude Opus 5.5 runs and handed to DeepSeek V4.1 Flash passed 58.3 percent of tasks on EnterpriseOps-Gym. That is also a vendor test.
Delivered is not used
The counter evidence arrived the same week. Delivered, Not Used, a preprint posted on 1 November, traced memory deliveries across 8,000 episodes on SWE-bench Verified from three memory systems. Resolve rates stayed within 0.4 percentage points of a no memory baseline. Most delivered entries were irrelevant. Between 51 percent and 80 percent of the relevant entries went unused, depending on the system, and in two systems most deliveries restated content the agent could still see. The paper also found that 23.2 percent of episodes with beneficial memory use still failed, while five of 16 episodes with harmful use still succeeded. A final score alone would have hidden both facts.
That finding changes what a compiled record has to prove. A wiki, a context set and a skill can all be well formed, current and approved, and still have no effect if the agent does not use them, or uses them on the wrong task. Compilation moves the cost earlier. It does not remove the need to check the result at the point of use. The earlier context retrieval test made the same point for retrieved passages: a passage is not a record until its source and scope can be inspected.
The Compilation Receipt Test
A compiled layer should pass three checks before a company trusts it as its record. First, source and quote: every compiled fact carries its source, its date, the version it superseded and a link or contiguous quote that lets a person check it against the original, as RAGFlow requires for a proposition and Zep requires for a skill step. Second, approval and rollback: promotion into the layer an agent treats as official goes through a named approver, is versioned, and can be rolled back, as Google provides for stewards and as Anthropic provides by leaving the input store untouched while a dream writes a separate output store for review. Third, a use receipt: the system logs which compiled items were delivered, which ones the agent actually used, and what followed, so a team can separate not found from found but not used, the distinction the Delivered, Not Used framework was built to make.
Those checks are the operating standard behind how Jilu works: a draft waits for review before it joins the shared record, and a published record keeps its source, date and superseded version visible. The context engineering pillar set the boundary for self editing context. Compilation extends that boundary. The compiled layer is derived data, and the original stays the source of truth.
Compiling knowledge ahead of time is now the direction of the market, from model labs to cloud platforms to open source retrieval. The gain is real when the compiled result is selective, checked against its source and measured at the point of use. Without those three receipts, a company has not built a better record. It has built a faster way to repeat a note it cannot verify, at a scale search never reached. That is the work persistent agent memory also has to pass: not whether the agent remembers, but whether what it remembers can be corrected, traced and shown to have been used well.