Skip to main content

How Memories Are Processed

When you store a memory, Memsolus automatically processes it in the background. You don't need to configure anything — the processing happens for every memory you add.

What Happens After You Store a Memory​

You store a memory
        ↓
AI processes & extracts key facts
        ↓
Deduplication check
        ↓
Entities extracted → Knowledge Graph
        ↓
Related memories consolidated → Knowledge Base
        ↓
Memory is READY and searchable ✓

Here is what each step does:

  1. Processes and normalizes the content — The text is analyzed and broken down into clear, atomic facts. If a single memory contains multiple distinct pieces of information, it may be split into separate memories automatically.

  2. Deduplicates against existing memories — Memsolus checks whether this information is already stored. If an identical or very similar memory exists, the new one is merged rather than stored as a duplicate.

  3. Extracts named entities — People, organizations, places, and other named concepts are identified and added to the Knowledge Graph, so you can query relationships later.

  4. Consolidates related memories — Similar memories are grouped and synthesized into structured knowledge entries, making it easy to retrieve a unified view of what the system knows about a user.

  5. Makes the memory searchable — Once processing is complete, the memory becomes available in search results via hybrid, semantic, or keyword search.

Memory Status​

The status field on a memory reflects where it is in the processing flow:

StatusWhat it means
PENDINGMemory received, queued for processing
PROCESSINGBeing analyzed by AI
READYFully processed and searchable
CONSOLIDATEDMerged into an existing memory that covered the same information

Processing typically completes within a few seconds. You can poll the memory's status or listen for the memory.processed webhook event.

Memory Layers​

Every memory has a layer that determines how it is treated:

Standard memories (default)​

Standard memories are permanent. They go through the full processing pipeline and remain until you explicitly delete them. This is the right layer for most information: user preferences, project facts, decisions, and any knowledge you want to persist long-term.

TASK memories​

TASK memories are short-lived. Use them for ephemeral context — things that are relevant only within a session or a specific task, such as the current step in a workflow or temporary working notes.

TASK memories:

  • Are processed more lightly than standard memories
  • Expire automatically based on your workspace settings
  • Can be promoted to a standard permanent memory if you decide the information is worth keeping
// Store a short-lived task memory
await client.memories.add({
  content: 'User is currently in step 3 of the onboarding flow.',
  userId: 'user_123',
  layer: 'TASK',
});

// Promote it to permanent when relevant
await client.memories.promote('mem_abc123');

Deduplication​

Memsolus detects duplicate memories automatically using two methods:

  • Exact match — If the same text is submitted twice, the second submission is discarded immediately.
  • Semantic similarity — If a new memory is very similar in meaning to an existing one, the two are compared and merged during processing. The result is a single, updated memory rather than two redundant entries.

This keeps your workspace clean without requiring you to manage duplicates manually.