Design a decay policy for an agent that has been talking to one user for two years

**Decay should not be a single dimension.** A score computed from one signal will always be wrong for some class of memory. The three signals to combine are: - **Recency**, last time the memory was accessed or referenced. - **Frequency**, how often it has been relevant over the recent past. - **Importance**, a flag set at write time for facts the model judges durable (safety, preferences, identity). Compute a salience score per memory periodically: a weighted sum of these three. Below a threshold, the memory is **demoted to cold storage**, not deleted. Cold-storage memories are still queryable by an explicit recall tool, but no longer eligible for default retrieval into the working context. This avoids the worst failure mode (losing a fact the user expected to still be remembered) while keeping the active store small. Two classes of memory must be **exempt from decay entirely**: safety-relevant facts (allergies, medical conditions, names) and explicit user-set preferences. These never decay regardless of recency. And **episodic memories** (specific dated events) should decay faster than **semantic memories** (distilled preferences) because dated facts are usually less re-relevant. Zep implements exactly this in production, with temporal weighting and importance flags. Mem0 leans more on deduplication and merging than on time-based decay; Letta gives the agent tools to manage its own memory rather than running a background decay process.

**The decay design starts from a refusal to use one signal.** Recency alone fails because a stable preference the user has held for two years and queries occasionally would get evicted. Frequency alone fails because something marked safety-critical may be queried rarely. Importance alone fails because old high-importance facts that the user has explicitly contradicted should still leave. The robust policy combines three orthogonal signals into a salience score: **recency** (timestamp of last access), **frequency** (access count over the last N days, decaying older accesses), and **importance** (flag set at write time, with an explicit class system for safety, preferences, and identity). Compute the score on a schedule, not per-read. The action below threshold is **demotion, not deletion**. Demoted memories move to cold storage where they remain queryable via an explicit recall tool but are no longer auto-injected into the working context. This is the single most important design choice, irreversible deletion creates user-facing failures that demotion does not. Differential decay by memory type matters. **Episodic memories** (specific dated events: 'last Tuesday the user mentioned Project Alpha') decay faster because their relevance window is naturally bounded. **Semantic memories** (distilled preferences: 'the user prefers concise answers, works in Python') decay slower because they describe stable properties. **Procedural memories** (how the agent and user typically interact) decay slowest of all. Finally, a hard-rule exemption list that never decays: safety facts (allergies, medical conditions, dependent names), explicit user-set preferences ('always use metric units'), and identity (name, primary contact info). These survive any decay sweep. Zep in production runs exactly this style of temporal weighting on its knowledge graph. Mem0 emphasizes deduplication. Letta exposes the policy to the agent itself.