Memory Manifolds
Memory Manifolds runs TRACE retrieval over the points currently loaded in a map. Instead of answering “which points are nearest the query”, it answers “which points does evidence flow to from the query”, by building a small graph around the query and letting a random walk spread across it. When the region around your query is a genuine cluster, that pulls in related material a plain nearest-neighbour list ranks too low. When it is not, the panel says so instead of inventing an order.
Every stage is shown rather than hidden: you can step through the probe, the local geometry, the graph, the walk, and the final ranking, and the map recolors to match whichever stage you are looking at.
Where to find it
View menu, Exploration section, Memory Manifolds.
The panel also declares a point action, Retrieve Along Manifold, for selections of one or more points. It is off by default: tick it under Selection Actions in the Settings panel and Memory Manifolds opens itself whenever you select points. The same action is listed under Analyze in the right-click menu on a selection in Map (Legacy), whether or not the checkbox is on.
Seeding a run
There are two ways to start, and both end in the same place: a query vector, which is the centroid of the seed points.
- Type a query and press Find. Mantis runs a semantic search over the space and seeds from the top five matches that are loaded in the current view. This needs an open connection to the space backend and a semantically indexed space; if matches come back but none are loaded, the panel tells you to zoom out so they load.
- Select points on the map and press Run from selection. Any selection works, from a single point to a lasso.
The line above the buttons reports what was used to seed and how many points are loaded. Clear discards the run and removes the panel’s coloring from the map.
Only loaded points take part
The pipeline reads the points the map has loaded, not the whole space. Level of detail means a zoomed-in view holds fewer points, which changes the candidate pool and therefore the result. The panel prints the loaded count so you can tell.
Mantis picks the richest vectors available: latent feature values when enough loaded points carry them, otherwise the 2-D map coordinates. The line under the stage stepper says which was used, along with the candidate count and how long the run took.
The five stages
Click a stage in the stepper to inspect it. The map recolors as you move between them, and clicking any candidate in the panel flies the map to that point.
1. Probe
A single cosine fetch of the nearest C candidates to the query. This is the ordinary vector-store lookup; everything afterwards re-reads only these candidates. The stage lists the cosine head and shows the distance of the nearest candidate and of the probe edge. On the map, candidates are tinted by nearness and the seed is green.
2. Pilot
The nearest k0 candidates are used to chart the local geometry. The chart is the eigen-spectrum of their covariance, and it reads left to right as “how much variation lies along the first direction, the second, and so on”.
- Intrinsic dim d̂ is the participation ratio of that spectrum: the effective number of directions that actually carry variance. A d̂ near 2 means the neighbourhood is essentially a surface, a large d̂ means it spreads through many directions at once, which is much harder to retrieve along.
- Conditioning κ compares the strongest direction to the weakest one still carrying signal. A very large κ means the local covariance is degenerate and the geometry should not be trusted; the verdict stage will abstain on it.
Pilot points are orange on the map.
3. Operator
A symmetric k-nearest-neighbour graph over the query and the candidates, with heat-kernel edge weights whose bandwidth is set per node from its own neighbourhood scale, so dense and sparse regions are treated on equal terms. The minimap draws the candidates at their true map positions with the graph’s edges.
- Nodes / edges size the graph.
- Mean bandwidth ε is the average local scale.
- Spectral gap λ1−λ2 measures how cleanly the graph separates into a well-connected community. A wide gap means there is one coherent clump for the walk to find; a gap near zero means the neighbourhood has no clear structure.
4. Walk
A personalized PageRank seeded at the query: mass starts at the query and floods along the graph’s edges, restarting at the query with probability α at each step. Press Play or drag the slider to scrub through the iterations. The panel reports the mass still sitting on the query at step t, and the map mirrors the minimap at every step.
The point of the walk is that mass reaches evidence connected through intermediates, which a straight-line distance would never rank highly.
5. Ranking
The final order, plus the panel’s verdict on whether to trust it.
Three orders are available from the toggle: diffusion (by final walk mass, cosine breaking ties), geodesic (by shortest path through the graph), and cosine (the plain probe order). Next to each row is its movement against the cosine order, so ↑7 means the manifold ranks that point seven places higher than similarity alone did. more / less switches between the top 15 and the top 30.
Reading the verdict
The banner at the top of the ranking stage is the router’s call on this specific query, taken from the same walk you just watched.
| Verdict | What it means for you |
|---|---|
| diffusion | Mass settled into a compact community that is not just the cosine head. This is the case Memory Manifolds exists for: use the diffusion order. |
| cosine | Either the mass collapsed onto the single nearest neighbour, or the walk simply reproduced the cosine head. Nothing was gained by the geometry, so read the plain similarity order. |
| abstain | Mass smeared across the whole pool, or the pilot covariance was too ill-conditioned. There is no evidence patch to assemble here, and the panel declines rather than dressing up a weak result. |
The numbers beneath it are the evidence for that call: mass concentration is the share of the walk’s mass held by the top few candidates, participation ratio and pr / pool say across how many candidates the mass is spread, peak mass is the largest single share, head overlap is how much the diffusion top matches the cosine top, and κ is the pilot conditioning again.
Operator settings
Open Operator settings to change the pipeline’s parameters. Each change takes effect on the next run.
| Setting | Range | Default | Effect |
|---|---|---|---|
| Candidates C | 50 to 600 | 250 | How many probe results enter the graph. Larger pools see more, and cost more. |
| Graph k | 4 to 20 | 10 | Neighbours per node. Lower keeps the graph sparse and local. |
| Pilot k0 | 10 to 60 | 30 | How many candidates chart the local geometry. |
| Walk steps t | 5 to 40 | 20 | Iterations of the diffusion walk. |
| Restart α | 0.05 to 0.5 | 0.15 | Probability of jumping back to the query. Higher keeps mass close to the query. |
Limits
- The run is computed in your browser over loaded points only. It is not a backend search and does not see the rest of the space.
- A run needs at least five usable points in the pool and seeds that have vectors in the same space as the pool.
- Results are not saved. Closing the panel or pressing Clear discards the run; there is no export.
- While a run is active the panel recolors points on the map. Clear restores their normal colors.
- Semantic seeding gives up after 20 seconds and reports why.