Definition
Memory lifecycle is the management of knowledge across its lifespan — from raw observation through consolidation to potential decay. Introduced by the LLM Wiki v2 extension, it addresses the problem that not all wiki content is equally valuable forever. Knowledge has a lifecycle, and the wiki should reflect that.
Key Points
- The original LLM Wiki treats all content as equally valid forever; v2 adds temporal awareness
- Four consolidation tiers: Working memory (recent observations) → Episodic memory (session summaries) → Semantic memory (cross-session facts) → Procedural memory (workflows and patterns)
- Each tier is more compressed, more confident, and longer-lived than the one below it
- The LLM promotes information up the tiers as evidence accumulates
- Ebbinghaus forgetting curve: retention decays exponentially with time; each reinforcement (access, confirmation from new source) resets the curve
Related Concepts
- llm-wiki — The original pattern that v2 extends with lifecycle management
- ingest — The operation that feeds information into the lifecycle
- lint — Detects stale claims and orphan content for lifecycle management
- agent-memory-systems — Broader field of persistent memory that includes lifecycle concepts
Add-only Growth and the Real Expiry Gap
A 2026-09-02 agent-memory deep-dive shows the practical version of the lifecycle problem in a shipped product. Mem0’s add-only extraction means nothing is ever deleted automatically; there is no built-in expiration or decay (open feature request #5330), and a year-old store just keeps accumulating stale facts — noisier retrieval plus rising storage/embedding costs. The documented workaround is manual: store short-lived facts with an explicit expiry in metadata and sweep or filter at read time.
This is the “who deletes stale truth” question the lifecycle theory poses, now observed in production. mem0 leaves expiry entirely to the operator; the from-scratch SQLite pattern sidesteps it by letting you overwrite or delete a row when a fact changes, the way you would handle any row in a database you own. Meanwhile memory-retrieval-patterns shows retrieval can weight recency as a signal, but that weighting is not reliable enough (issue #4956) to substitute for an explicit expiry/decay policy.
Implications
For long-running personal memory, add-only durability and lifecycle forgetting pull in opposite directions: preserving history and removing stale truth both matter, and a robust memory system needs an explicit decay/expiry mechanism rather than assuming retrieval ranking will bury outdated facts.
Implications
The memory lifecycle turns the wiki from a flat collection of equally-weighted claims into a living model where the LLM can express certainty. It prevents the wiki from becoming a “junk drawer” of equally old and equally stagnant content. For this vault, implementing lifecycle would mean adding confidence scores to claims and a mechanism for supersession when sources contradict.
Open Questions
- Is numeric confidence scoring (0.85) false precision, as critics argue? The real signal may be the chain of links a claim carries, not a float.
- Should forgetting be automatic or human-determined? Critics argue that old doesn’t mean stale — a bug from six months ago is often more valuable than one from last week.
- How does the vault distinguish “never accessed” from “not useful”? Access patterns may not correlate with value.