Memory-as-Metabolism: Digestion-Based Agent Architecture

agent_domain: Customer support agent for a SaaS product, memory_budget: 32,000 tokens per session with 2,000 token persistent store, retention_priority: user frustration signals and unresolved issues, interaction_frequency: ~15 conversations per day per user account

## 1. INGESTION LAYER Information enters via three channels: real-time user utterances (streamed token-by-token), retrieved knowledge base snippets (batch, pre-session), and internal monologue traces (async). A circular buffer holds the last 8 raw turns before enzymatic processing begins. Ingestion triggers fire on: sentence boundary, topic shift signal (cosine distance > 0.4 from running centroid), or explicit user emphasis markers ('importantly', 'always', 'never'). ## 2. ENZYMATIC DECOMPOSITION Four enzymes run in parallel on each ingested chunk: - **Temporal Stripper**: Removes tense and date anchors, converting 'my subscription expired last Tuesday' → [subscription_status: expired, recency: high] - **Frustration Extractor**: Scores emotional valence and maps to frustration taxonomy (billing_friction, feature_gap, response_latency) - **Entity Normalizer**: Canonicalizes product names, plan tiers, and account IDs against a lookup table - **Resolution Classifier**: Tags each unit as resolved / unresolved / ambiguous based on conversational closure signals ## 5. ELIMINATION PROTOCOL Memory units carry a half-life determined by class: resolved issues decay with t½ = 2 sessions; billing complaints carry t½ = 14 days (high liability); feature requests are permanent until shipped. Elimination triggers include: explicit user correction ('never mind, that's fixed now'), contradictory signal (subscription shows active after 'expired' was absorbed), and a nightly sweep that expels any unit scoring below 0.1 on a composite salience × recency function. Crucially, eliminated units are not deleted — they are composted into aggregate statistics (e.g., 'billing friction mentioned 3× in last 30 days') before purge, preserving metabolic byproducts. ## 7. CONTRAST WITH RAG | Dimension | MaM | RAG | |---|---|---| | Latency | Higher upfront, faster retrieval | Low upfront, retrieval-dependent | | Adaptability | Self-updating belief state | Static until re-indexed | | Forgetting | Active, rule-governed | Passive (retention by default) | | Synthesis | Generates new inferences | Returns source chunks | | Interpretability | Traceable absorption log | Opaque vector similarity |