# Neural Nexus
**Neural Nexus is a universal, self-hosted memory system that gives AI agents persistent, structured, human-owned memory across conversations, applications, models, and providers.**
Most AI assistants begin each conversation with limited knowledge of what happened before. Neural Nexus provides a shared memory layer where authorized agents can capture interactions, recall relevant information, track changes over time, and preserve the history behind what they know.
## Persistent Memory Across AI Systems
Neural Nexus is designed to work independently of any single AI model or provider. OpenAI-compatible applications, Ollama models, local models, MCP clients, custom agents, and other host applications can all use the same memory system.
This allows a person's information to remain available when they:
- Switch between AI models or providers
- Use multiple agents on the same project
- Move between local and cloud-based models
- Continue work in a later conversation
- Replace an application without losing its accumulated context
The human user owns the memory space. Agents are recorded as the source of actions and information, but they do not become the owner of the user's memories.
## Automatic Conversation Capture
Neural Nexus can capture ordinary conversations through supported proxy and host-adapter integrations. The host application forwards visible events to Neural Nexus as they occur, including:
- User messages
- Assistant responses
- System messages
- Tool calls and tool results
- Command executions
- File operations
- Errors
- Task-state changes
- Completion outcomes
Each captured event is preserved in a session-level episodic record and also processed into durable memory. This allows Neural Nexus to maintain both the original conversational history and a structured long-term representation.
MCP remains available for explicit memory operations and as a fallback integration method. Because MCP alone cannot observe conversations when an agent does not call a tool, automatic capture requires a proxy, host adapter, or provider-specific integration.
## Capture-First Intake
Every interaction received by Neural Nexus is captured as a memory. Intake does not decide whether the information deserves to exist.
The intake system classifies and annotates each captured memory so it can be organized and retrieved effectively. A local language model, such as `llama3.2:1b`, can assist with semantic interpretation by selecting from controlled choices such as:
- Fact
- Entity
- Preference
- Decision
- Rule
- Subject type
- Preference polarity
- Preference strength
- Content type
Model responses are validated against a closed schema. Invalid responses are returned to the model for correction. If correction attempts still fail, deterministic application logic supplies safe defaults and the memory remains captured.
Storage, ownership, authorization, hierarchy, retention, and lifecycle decisions are enforced by Neural Nexus code rather than delegated to the language model.
## Structured Organization
Neural Nexus organizes memories into a consistent hierarchy:
- **Drive:** The broad person, project, team, system, or area
- **Partition:** The kind of memory, such as facts, preferences, decisions, rules, or events
- **Folder:** The specific subject within that partition
- **Links:** Connections to related people, projects, topics, and memories
This structure works alongside semantic search. Users can search broadly by meaning or narrow retrieval to a particular project, category, folder, time period, or linked subject.
## Intelligent Recall
Neural Nexus combines multiple retrieval methods rather than relying entirely on vector similarity.
Recall can incorporate:
- Semantic similarity
- Keyword and exact-term matching
- Titles
- Entities and aliases
- Hierarchy fields
- Memory links
- Temporal relevance
- Current or historical status
- Provenance
- Memory strength
- Reinforcement history
- Controlled recency
- Second-stage reranking
Candidate results are fused, reranked, deduplicated, and selected according to their evidence value and token cost. This helps retrieve precise context without filling an AI model's context window with loosely related material.
## Time, Changes, and Contradictions
Information changes. Neural Nexus represents those changes without silently erasing the past.
Memories can include:
- When information was observed
- When it became valid
- When it stopped being valid
- Whether it is current, historical, disputed, or superseded
- Which memory it replaced
- Which memory later replaced it
- The source events supporting it
- Confidence and relationship metadata
When a preference, decision, or fact changes, Neural Nexus can retain the earlier statement while identifying the newer statement as current. Historical queries can retrieve the information that was valid at an earlier time.
Conflicting evidence can remain visible instead of being forced into a false conclu