Knowledge lifecycle management¶
Definition¶
Knowledge lifecycle management is the practice of treating every claim in an accumulating knowledge base as having a lifecycle rather than being equally valid forever: claims carry confidence scores derived from supporting-source count, recency of confirmation and known contradictions; newer information explicitly supersedes older claims instead of sitting beside them; unreinforced material fades in priority along a forgetting curve; and observations consolidate upward through tiers — working, episodic, semantic, procedural — becoming more compressed, more confident and longer-lived at each step.
Explanation¶
Confidence scoring turns a flat pile of equally weighted assertions into a model that can grade its own certainty: a claim like 'Project X uses Redis for caching' knows it came from two sources, was last confirmed three weeks ago, and sits at 0.85 — confidence decays with time and strengthens with each reinforcement. Supersession is version control for knowledge rather than files: when new information contradicts an existing claim, the replacement is linked and timestamped to the original, which is preserved but marked stale, so contradictions resolve instead of accumulating as marginal notes. Forgetting keeps the base from becoming noisy: retention decays exponentially and every access or confirming source resets the curve, with the decay rate depending on claim class — architecture decisions fade slowly, transient bugs fast — and faded claims are deprioritized like a bottom drawer, never deleted. Consolidation tiers promote information as evidence accumulates: raw observations compress into session summaries, cross-session facts consolidate from episodes, and repeated patterns extract into procedures, which is how 'I saw this once' becomes 'this is how things work'. The source is a practitioner extension gist by the author of agentmemory, a memory engine for coding agents, reporting what broke when running the LLM-wiki pattern across thousands of sessions — an experience report that also promotes the author's project, not a measured study.
Key Properties¶
- Per-claim confidence from source count, confirmation recency and contradictions; decays with time, strengthens with reinforcement
- Supersession links and timestamps a replacing claim to the replaced one, preserving the stale version
- Decay is class-dependent — architecture decisions fade slowly, transient bugs fast — and faded claims are deprioritized, never deleted
- Consolidation tiers (working, episodic, semantic, procedural) are each more compressed, more confident and longer-lived than the tier below
Relationships¶
- Ebbinghaus forgetting curve — borrows its decay model for machine memory — retention falls exponentially with time and every access or newly confirming source resets the curve, with the decay constant varying by claim class
- LLM-maintained wiki — adds the layer that pattern omits: the base wiki treats all content as equally valid forever, whereas lifecycle management grades, supersedes and fades claims as evidence shifts
- Fitted spaced-repetition scheduling — knowledge lifecycle management applies the same decay-driven scheduling logic fitted spaced-repetition scheduling fits to a human learner's recall, instead to a knowledge base's claims — unreinforced material fading along a forgetting curve is the system-directed analogue of a personal forgetting estimate.
Applications¶
Keeping a long-lived agent memory or team wiki trustworthy as it grows: rank answers by claim confidence, resolve contradictions by supersession instead of accumulation, and let stale operational detail fade while durable decisions persist.
Sources¶
- https://gist.github.com/rohitg00/2067ab416f7bbe447c1977edaaa681e2