Persistent Knowledge Base

Title: Persistent knowledge base Slug: persistent-knowledge-base

A persistent knowledge base, in the context of agentic AI systems, is a structured repository of accumulated insights (failure patterns, successful strategies, rejected proposals, evolution history) that an AI agent's skill-evolution loop maintains across iterations and that, once written, is never reset. It differs from an inference-time retrieval-augmented generation index or a conversational memory buffer. The defining properties are persistence (survives across skill updates and rollbacks), accumulation (grows monotonically through deliberate consolidation rather than being overwritten), and separation from inference (the agent executing tasks does not read the wiki directly during rollout; instead, the skill proposer reads it to decide what skills to revise).

The concept crystallized in the WikiSkill framework from Google Research in August 2026, where it was formalized as the Wiki Layer between immutable raw execution traces and the evolving Skill Layer. The architecture is a response to the failure mode of earlier skill-evolution methods, in which each iteration started with a clean slate and the proposer had no structured context for what previous iterations had already discovered. Without persistence, the proposer repeatedly re-litigates the same failure modes, and rejected proposals are lost to future iterations. The wiki's role is to record what is believed to be true and what was tried and failed, so the next iteration can build on a record of the search rather than re-conducting it.

The technical design has three components. First, the Wiki Maintainer, an LLM-based process that analyzes each iteration's execution traces alongside the existing wiki and proposes additions or revisions, including new failure patterns, refined strategies, and corrected diagnostics. Second, the wiki's content schema, which typically includes a pattern catalog (what kinds of failure modes recur), an evolution log (what skills have been proposed and why), and a structured-knowledge section (consolidated strategies and corrections). Third, the access policy that determines who can read and write the wiki during which phase of the loop. WikiSkill's ablation showed that giving the Skill Proposer wiki access produced a 15-percentage-point gain, while giving the Inference Agent direct wiki access during training rollouts degraded performance. The persistence is for the editor, not the executor.

The persistent-knowledge-base pattern generalizes beyond agent skill evolution. Any system that iteratively refines an artifact, such as prompt templates, code generation policies, search query reformulators, or planning heuristics, faces the same structural question of whether to remember past iterations or reset. The wiki pattern is the persistent alternative: a structured, append-mostly memory that accumulates diagnostic context across attempts and that the refining component reads to decide what to change next. This makes persistent knowledge bases a candidate architectural primitive for any self-improving agent whose improvement loop runs more than once.

The closest neighboring concepts are agent memory architectures more generally, including long-term memory modules, episodic buffers, and retrieval-augmented generation stores, but the persistent knowledge base is narrower. It is specifically about knowledge accumulated during the system's own evolution, indexed for the proposer, and never serving as the agent's direct retrieval source during task execution. This separation is what allows the wiki to grow without paying an inference-time context cost and what keeps failed experiments from silently degrading future runs.