What is Knowledge consolidation?
Folding scattered findings from past work into maintained articles, and correcting them when the facts change.
Memory without curation becomes a pile; curated knowledge is what the next task can actually use.
On this page
Knowledge consolidation is the ongoing work of turning scattered findings from completed tasks into a small set of maintained, trustworthy articles — merging what repeats, correcting what changed, retiring what is no longer true. It is the difference between an agent workspace that remembers and one that merely accumulates: transcripts, logs and comments are memory in the geological sense — everything is in there and nothing can be found.
The need follows from the economics of agent work. A run that discovers something expensive — an edge case, a failed approach, a corrected assumption — produces that finding once. Without consolidation, every future run pays to rediscover it or, worse, acts on a contradictory half-record of it. With consolidation, the discovery made on Tuesday is a summary line for every run on Wednesday.
What consolidation actually does
Consolidation is four operations, and only one of them is writing.
| Operation | What it means | What it prevents |
|---|---|---|
| Fold | A finding lands in the article that owns its topic | Topic sprawl; the same lesson in five places |
| Merge | Overlapping records become one article with one claim | Contradictions retrieved together — the agent gets a coin flip where it needed an answer |
| Correct | A stored claim that turns out wrong is edited, with the correction versioned | Acting on known-wrong premises; silent overwrites of what people relied on |
| Retire | Claims no longer true — or never verifiable — leave the live set with a record | Stale knowledge arriving with the system's authority behind it |
The correction and retirement rows are where most systems are weakest, and where consolidation earns its name. Appending is easy; every note-taking tool does it. Consolidation means the store converges on truth instead of growing toward noise.
Why retrieval alone does not solve it
Vector search over an unconsolidated store returns matches, not answers. The store holds the conclusion next to the dead ends that preceded it, next to the version of the claim that was later corrected, next to the duplicate with slightly different wording — and retrieval ranks them all with serene indifference to which is true. The context-rot findings apply to everything retrieved: whatever enters the window competes for attention, true or not.
That is why consolidation and retrieval are complements, not alternatives. Retrieval decides what the agent sees; consolidation decides whether what it sees is worth seeing. Curated, versioned articles with summaries injected first — full text opened only when relevant — are the shape that survives both requirements; see AI agent memory for the full mechanism.
How the work should happen
Consolidation fails when it is scheduled as its own activity — a monthly "clean up the knowledge base" day that never arrives. It works when it is attached to the flow that produces the findings:
- At the end of a run, anything new is folded into the article that owns the topic, or starts a new one if no owner exists.
- During a change, when the work invalidates a stored premise, updating the article is part of completing the change — the living documentation discipline.
- On contradiction, when a new finding disagrees with a stored claim, the store is reconciled then and there: one of them is wrong, and both being retrievable is the bug.
Two quality rules keep the store healthy. Every stored claim should earn its space — "which future agent, doing what, needs this?" — and size caps should force consolidation by making hoarding impossible: an article that grows without bound becomes a transcript with a title.
Knowledge consolidation in YATFA
Consolidation is part of the loop in YATFA, not a chore beside it: findings from completed work fold into versioned knowledge articles that every agent in the project shares, corrections land as versions with the history preserved, retirement is a deliberate, recorded operation, and summaries inject into new runs with the full text opened only on demand. See living knowledge and the design write-up in memory for AI coding agents.
Frequently asked questions
How is consolidation different from summarisation?
Summarisation shortens one thing; consolidation reconciles many. A summary of a run is useful; consolidation is what decides whether it becomes a new article, a correction to an existing one, a merge, or nothing because the store already knows it. The reconciliation is the hard and valuable part.
Who does the consolidating?
In an agent platform, the agents do — at the end of runs and during changes, when the facts are in context and the correction costs a paragraph instead of an investigation. Periodic sweeps catch what the flow missed. Human review stays available for high-impact corrections, under the project's autonomy settings.
How do you know consolidation is working?
Two readable signals: reuse (runs that find their answer in memory instead of re-deriving it) and contradiction rate (how often retrieval surfaces two claims that disagree). A healthy store shows rising reuse and near-zero contradictions; a growing store with flat reuse is a pile, not a knowledge base.