Advanced Research & Agentic Memory¶

Embenx implements several state-of-the-art algorithms and specialized memory structures to handle the complex requirements of autonomous agents and high-throughput retrieval.

Research-Driven Optimizations¶

Embenx stays at the cutting edge by integrating recently published research directly into the toolkit.

ClusterKV (arXiv:2412.03213)¶

ClusterKV improves retrieval throughput by semantic grouping. Instead of searching the entire index, Embenx can identify the most relevant semantic “neighborhood” and focus its compute there.

from embenx.core import ClusterCollection

col = ClusterCollection(n_clusters=10, dimension=768)
col.add(vectors, metadata)
col.cluster_data()  # Computes K-Means centroids

# Searches within the nearest cluster for speed
results = col.search_clustered(query_vector)

TurboQuant (arXiv:2504.19874)¶

TurboQuant addresses the “Memory Wall” in LLM serving by aggressively compressing activation tensors (KV cache). Embenx uses a sign-based 1-bit quantization scheme to reduce activation storage by up to 4x while maintaining high signal.

# Activations are automatically packed into int8 safe-tensors
col.add_cache(vectors, activations, metadata, quantize=True)

Agentic Memory structures¶

Autonomous agents require memory that is self-healing, time-aware, and spatially grounded.

Self-Healing Retrieval¶

The AgenticCollection allows agents to provide feedback on retrieval quality, which is then used to bias future results.

Positive Feedback: Decreases the semantic distance of a document for future queries.
Negative Feedback: Increases the distance, effectively demoting “noise” or irrelevant results.

col = AgenticCollection(dimension=768)
col.feedback(doc_id="doc_abc", label="good")
results = col.agentic_search(query_vector)

Temporal Memory (Echo, arXiv:2502.16090)¶

Agents often need to remember “what happened when.” The TemporalCollection supports recency-biased retrieval and time-window filtering.

col = TemporalCollection(dimension=768)
col.add_temporal(vectors, timestamps=my_unix_timestamps)

# recency_weight=0.7 gives high priority to recent events
results = col.search_temporal(query_vector, recency_weight=0.7)

Spatial Memory (ESWM, ICLR 2026)¶

Inspired by neuroscience, Episodic Spatial World Memory (ESWM) grounds embeddings in physical coordinates. This is essential for embodied agents (robots, drones) that need to retrieve memories based on their current location.

col = SpatialCollection(dimension=768)
col.add_spatial(vectors, coords=xyz_positions)

# Find memories near the agent
results = col.search_spatial(query_vector, current_coords=[0, 0, 0], spatial_radius=10.0)

Managed Sessions¶

The Session class provides a high-level manager for agentic memory, handling persistence and temporal decay automatically. Each session is stored as a dedicated Parquet file, making multi-user agent systems easy to scale.

from embenx.core import Session

sess = Session(session_id="agent_alpha", dimension=768)
sess.add_interaction(vector, "Agent thought or observation")
context = sess.retrieve_context(query_vec)