← Technology OverviewLiving Memory — The Substrate

Not a database.
A gravitational
system.

Every memory has a gravity score. High-gravity memories orbit close to their cluster. Low-gravity memories drift to the periphery. The nightly engine recalculates everything.

What you see in the background is a live simulation of the substrate topology — clusters, memories, and the relationships binding them.

Knowledge Cluster
Entity Cluster
Memory record
01 — Signal Ingestion

What enters the substrate.
How it gets there.

Every signal — real-time chat, imported history, agent findings, corrections — passes through the same five-stage ingestion pipeline before being written to the substrate. No exceptions. No shortcuts.

Raw Signal
Chat · Document · Agent · Correction · Import
Entity Resolution
Named entities extracted. Aliases mapped. Co-occurrence inferred.
Vector Embedding
Semantic embedding generated. Topic similarity axis anchored.
Gravity Scored
Initial weight calculated. Emotional density, frequency, recency.
Substrate Write
Written to PostgreSQL. Cluster membership assigned or created.
02 — Four-Axis Recall

Entity. Temporal. Well. Topic.
All four. Simultaneously.

Retrieval runs across four independent dimensions — none of which is a standard vector search. The five-axis scoring engine in the kernel uses these four dimensions as inputs. This is where “compounding” becomes structural, not metaphorical.

01

Entity Axis

Who

Recall surfaces everything connected to a specific entity — people, organizations, projects, concepts. Alias resolution ensures the graph is complete, not fragmented by name variation.

02

Temporal Axis

When

Time-scoped retrieval. Recent context weighted differently from longitudinal. The system knows the difference between a decision made last week and a pattern established over six months.

03

Well Axis

Depth

The Gravity Well score — a composite of frequency, recency, emotional density, and outcome correlation. Deep wells surface first. Shallow wells only when directly relevant.

04

Topic Axis

What

Semantic topic clustering. Vector similarity anchors retrieval to a concept space, then structural and relational signals amplify the most operationally significant results within that space.

03 — Knowledge Clusters

Memories organize.
Clusters compound.

Individual memories are assigned to Knowledge Clusters — topic groupings with measured cohesion, leakage, drift, and salience. A cluster isn't a folder. It's a living structure with health metrics, recalculated every night by the engine.

Cluster
Memory
Entity
Arc
Decision
Pattern
Topic

Cohesion

0.84

Average membership strength across all nodes in the cluster. High cohesion = tight, operationally specific topic.

Leakage

4.2%

Percentage of cluster members below the gravity threshold. Managed nightly — leaking members pruned or reassigned.

Drift

0.11

Rate at which the cluster centroid is shifting topic-wise. High drift triggers canonical summary regeneration.

Salience

0.91

Composite score used by five-axis retrieval. Determines whether the cluster amplifies or attenuates individual memory scores.

04 — Gravity Weighting

Not all memory is equal.
Gravity decides what surfaces.

Every memory has a gravity score — a composite of frequency, recency, emotional density, and outcome correlation. High-gravity memories surface first in retrieval. Low-gravity memories require direct relevance before they appear.

The nightly engine recalculates total_gravity and active_gravity for every memory in every cluster that drifted more than 0.5 since the last cycle.

Gravity score examples — same substrate

Q3 budget decision, revisited 12x
0.94
Key vendor risk — flagged, corrected
0.87
Communication preference pattern
0.81
New topic — 2 mentions, no outcome
0.31
Offhand comment, never revisited
0.12
05 — Correction Propagation

One flag.
Immediate propagation.

When a user marks something incorrect, the correction doesn't just suppress a single memory — it propagates through the graph. Gravity is adjusted. Context flags are set. Affected clusters are queued for recalculation. Five steps. Real-time.

Flag Received
User marks incorrect
Gravity Reduced
Memory downweighted immediately
Context Flag Set
Override marker written
Cluster Recalculated
Affected clusters re-examined
Propagation Complete
Next retrieval reflects correction
Next

See the operating layer that runs on top of this.