Retrieval Pipeline

Multi-signal search. One answer.

Every recall runs multiple search strategies in parallel across your memory store and knowledge graph. Our proprietary scoring algorithm combines results into a single ranked list.

The Pipeline

1

Understand

Analyses the query to determine what kind of memory retrieval is needed

2

Optimise

Rewrites and expands the query for maximum retrieval accuracy

3

Search

Runs multiple search strategies in parallel across your knowledge graph

4

Score

Proprietary fusion algorithm combines results from all search paths.

5

Assemble

Deduplicates, formats, and trims to your token budget. Quality over quantity

Why It Works

Hybrid Graph-Retrieval

Combines knowledge graph traversal with semantic search. Finds connections between entities that pure embedding similarity misses entirely.

Sub-200ms

Multi-Signal Search

Multiple search strategies run in parallel. Each catches different types of results. The fusion layer combines them into a single ranked list.

Parallel execution

Temporal Awareness

Understands time. When you ask 'what happened last week', the pipeline knows to search by date range, not just semantic similarity.

Built-in

Knowledge Updates

When facts change, old information gets archived and new information takes priority. No stale results polluting your context window.

Automatic

The Result

83.6% accuracy on the LongMemEval-S benchmark, running gpt-4o-mini. Ahead of the industry leader best system on the leaderboard, which uses the significantly more expensive GPT-4o. The retrieval pipeline does the heavy lifting, not the model. That means better results at a fraction of the cost.

524

tokens

Median context

688

tokens

Average context

1,642

tokens

95th percentile

Selective retrieval by design. Only what matters, not the entire context window.