MemX — Adaptive Memory Engine for AI
Zero-LLM-cost AI memory engine · bio-inspired forgetting curve · 12 PII filters
What is MemX?
MemX is Lurus’s adaptive memory engine for AI, built on ACE (Adaptive Context Engine)ACEAdaptive Context Engine. The core intelligence layer of MemX, comprising four modules: knowledge distillation, deduplication, decay, and retrieval.Learn more → v2.0. It provides AI agents with full knowledge lifecycle management: intelligent distillationKnowledge DistillationExtracting refined, reusable knowledge entries from large volumes of conversation. MemX's Reflector engine does this automatically. → semantic deduplicationSemantic DedupSemantic deduplication, which uses cosine similarity to automatically identify and merge knowledge entries that mean the same thing but are worded differently. → decay-based forgetting → hybrid retrieval, giving AI a truly human-like "memory."
Three core advantages
- Hybrid mode by default + automatic fallback — when the LLM is unavailable, it switches to pure rules: zero calls, zero cost.
- Bio-inspired forgetting curve — Ebbinghaus exponential decay, half-life defaults to 30 days; strongly recalled items are promoted to permanent memory.
- End-to-end privacy protection — sensitive information never enters the vector database.
Core Features
From conversation to retrievable memory
For detailed rules, formulas, and parameters, see Core Concepts and Architecture.
Intelligent knowledge distillation (Reflector)
Hybrid mode (rule-based prefiltering + LLM refinement) identifies 5 knowledge patterns: error fixes / retry successes / config changes / new tool usage / repeated operations, scoring each 0-100 to filter out low-score noise.
Semantic deduplication and conflict detection (Curator)
Cosine similarity ≥0.8 auto-merges, 0.5-0.8 flags potential conflicts, below 0.5 is treated as independent knowledge.
Bio-inspired memory decay
7-day protection period + exponential decay + recall reinforcement; items retrieved more than 15 times are promoted to permanent memory and no longer decay.
Four-layer hybrid retrieval
L1 exact → L2 fuzzy → L3 metadata → L4 vector; after ScoreMerger weighted fusion, multiplied by DecayWeight × RecencyBoost × ScopeBoost, with automatic fallback when the vector layer is unavailable.
Privacy-first design
12 built-in sensitive-information filter rules (keys / tokens / database connection strings / local paths / custom regexes), automatically blocked before write.
Architecture Overview
Knowledge flows in from conversations and passes in turn through distillation, privacy filtering, and deduplication before landing in vector and metadata storage; retrieval requests go through the four-layer hybrid pipeline, while the decay engine continuously maintains memory activity in the background.
Use Cases
Remembers your coding habits, the pitfalls you’ve hit, and project conventions.
Accumulates knowledge of customer interaction history to deliver personalized service.
Automatically distills and organizes knowledge from everyday conversations.
Shares team-level memory so new members can quickly get up to speed on context.
Comparison with Traditional Memory Systems
Why this is not just another vector store
| 维度 | MemX (ACE) | Traditional approach (mem0) |
|---|---|---|
| Knowledge extraction | Hybrid engine (rule-based prefiltering + LLM refinement, reducing calls by 90%+) | LLM (2-5K tokens each time) |
| Deduplication | Cosine-similarity auto-merge | LLM judges item by item |
| Forgetting | Exponential decay + recall reinforcement | Permanent storage, no eviction |
| Search | Four-layer hybrid search | Vector search only |
| Privacy | 12 built-in sensitive-information filter rules | No built-in protection |
| Scope | Hierarchical (global / project / workspace) | Flat (user / agent) |
| Token management | Built-in budget trimming (CJK-aware) | Caller manages it themselves |
| Local embeddings | ONNX local inference, fully offline | Requires an API |