Why Agent Memory Goes Beyond Vector Databases: Inside MNEMOVELA’s Cognition Runtime Model

Many agent systems implement “long‑term memory” with a simple pipeline: they save chat logs, split the text, generate embeddings, and later retrieve similar fragments from a vector database. This approach improves information recall but does not create true semantic memory.

Vector databases can answer “which contents are semantically similar,” yet they cannot independently answer critical questions such as:

Is this a raw experience or an inferred fact?

Is the fact still valid now?

How should conflicting records be handled?

Is the content from a real observation or a hypothetical simulation?

Which agent, tenant, project, or reasoning branch does it belong to?

How should retrieved results be assembled into the current context?

MNEMOVELA defines itself as a Cognition Runtime that sits beside large models and is responsible for long‑term memory, context assembly, and knowledge expression. Its data model distinguishes several memory types:

Episode : raw events such as what the user said and what the system returned.

Fact : subject‑predicate‑object statements with an expiration time.

Knowledge : structured cognition like entity relationships and classifications.

Experience : reflections formed after actions.

Simulation : hypothetical scenarios used for reasoning about possible outcomes.

Belief, Intention, Procedure, Mission, Preference : internal states and behavioral tendencies of the agent.

Distinguishing these types is essential. For example, a raw Episode should not automatically become a Fact just because it is retrieved; a Simulation should not be treated as real memory; Experience reflects reflection, not the action itself. Each type carries different credibility, validity, and usage rules.

MNEMOVELA also introduces a Git‑like branching model. An agent can create a branch from a particular memory state, perform simulations, experiments, or revisions within that branch, and avoid contaminating the main cognition line. Every record is immutable and appended‑only, carrying metadata such as branch, timestamp, retention level, tenant, and project scope. When a Fact is corrected, the system does not overwrite history; it appends a new version or relationship, preserving the evolution of cognition.

The retrieval layer uses a hybrid strategy that combines multiple dimensions:

Lexical matching

Vector semantic similarity

Entity and relationship matching

Temporal windows and validity periods

Branch and identity scope

After retrieval, a Context Assembly step reorganizes the candidate records based on the current task, token budget, memory type, and relevance, forming a context that the model can actually consume instead of merely returning the top‑scoring text chunks.

Public repositories provide Python, TypeScript, and Go clients, along with JSON‑RPC, MCP, REST, and gRPC contracts. The open materials also showcase a Web Console and a Coding Agent Continuity scenario that continuously use these interfaces.

The core thesis of MNEMOVELA is: “Memory is not a storage location; it is a set of semantic constraints about type, time, source, identity, branch, and usage.” While vector databases can serve as a storage or retrieval mechanism, they are not equivalent to Agent Memory. A genuine long‑term memory infrastructure must manage how records are generated, when they are valid, how they are revised, and in what context they may be reused.

Project repository: https://github.com/axisrobo/mnemovela-open