Working memory

L0: WORKING MEMORY  (Redis, TTL 1–60min)
    Tool outputs awaiting processing
    Intermediate computation results
    "Currently looking at file X" state
    Multi-step workflow checkpoints
    Rate-limit counters / dedup bloom filters
    Key: mnemo:{ws}:work:{agent}:{key}
    Eviction: TTL-based (never LRU)

L1: SESSION MEMORY  (Redis, TTL 24h)
    Session buffer (turns + pins + scratch)
    Conversation-scoped entity co-reference
    In-flight task state
    Key: mnemo:{ws}:sess:{id}:*
    Eviction: TTL + collapse trigger

L2: HOT CACHE       (Redis, TTL 5min–24h)
    ACT-R activation scores (sorted set, top-1000/ws)
    Recent recall results (query hash → hits)
    Agent profile summary
    RBAC policy cache
    Key: mnemo:{ws}:hot:*
    Eviction: LRU by activation score

L3: SEMANTIC CACHE  (PostgreSQL, TTL 24h)
    mnemo_query_cache (cosine ≥0.95 shortcuts full pipeline)
    Embedding LRU (avoid re-embedding same text)

L4: LONG-TERM       (PostgreSQL, bitemporal)
    All 10 cognitive primitives
    Knowledge graph (hypergraph relations)
    Reasoning library
    Citations + provenance

L5: ARCHIVE         (PostgreSQL, partitioned, immutable)
    mnemo_fact_archive (merged/pruned facts)
    mnemo_audit_log (Merkle-chained)
    mnemo_session_log (collapsed sessions)

Redis healthy     → Full stack (L0–L5)
Redis degraded    → L0/L1 fall to in-process Maps (bounded 10MB)
                    L2 served from PG (slower but correct)
Redis unavailable → Sessions disabled (503 on session ops)
                    Recall works (L3+L4 only, ~100ms not ~5ms)
                    Tasks work (PG-backed, no hot cache)