Choose a memory mode

Use competitive mode for an adaptive associative memory. Use plain mode when each supplied vector should remain a separate, unchanged location. Both modes support associative recall; they differ in how writes represent observations. The mode applies to every collection in a database.

Create an adaptive memory

curl -u admin:pw -X POST http://localhost:6380/db \
  -H 'Content-Type: application/json' \
  -d '{"name":"knowledge","dimension":128,"eam":{"competitive":true}}'

A novel write allocates a location without accumulating its content into existing locations. A joining write updates the activated representation. This supports both recall of distinct patterns and learning from repeated, related examples.

Use partial-cue reconstruction to measure recall and independent held-out examples to measure shared-pattern learning. Choose read settings using those measurements. Competitive mode does not guarantee exact reconstruction of every observation, especially when related observations share a representation.

When not to use learning

Choose competitive: false when your application needs each observation to stay as a separate vector location:

• Keep individual examples or episodes without combining related observations.
• Supply fixed address/value pairs and use attention to weight those entries.
• Add new observations without adapting the locations of earlier observations.

Create a plain database using the request below. Plan storage for a growing location count, and measure read latency as it grows. Repeated writes also add locations; plain mode does not automatically consolidate them. Explicit merge operations can change that representation.

Use competitive mode when repeated, related observations should update a shared representation or adapt it to changing patterns. Do not choose plain mode only to obtain document IDs: exact document lookup works with either mode.

The write-mode benchmark compares write costs with synchronous durability. Turning learning off skips competitive selection and adaptation; it does not skip persistence or make writes free.

Create a memory with separate vector locations

curl -u admin:pw -X POST http://localhost:6380/db \
  -H 'Content-Type: application/json' \
  -d '{"name":"episodes","dimension":128,"eam":{"competitive":false}}'

Plain mode preserves the supplied vectors as individual locations rather than learning a shared representation at write time. Attention still weights and blends locations: preserving a vector does not guarantee that every damaged cue will recover it exactly. Measure cue damage, occupancy and attention scale.

Choose the result you need

An associative read returns a reconstructed vector. It may match an individual pattern or represent shared structure; it is not a document identity.

Documents written with metadata retain document records independently of the learned locations. Use document IDs for exact ID/metadata lookup or documents/query for similarity-ranked document results. HTTP ID lookup returns the ID and metadata, not the vector. These APIs work with either write mode.

Separate competitive and plain databases are useful when you want both adaptive vector locations and an independent non-adaptive vector memory. A second database is not required merely to retain document records. Neither policy promises unlimited capacity or automatic unknown-query rejection.

competitive is selected at database creation, not per request. To choose a different representation, create a database and replay source observations. Changing a setting cannot reconstruct observations already combined in counters. Start with default learning settings; k=1 is an experimental alternative, not a general recommendation. Changing k affects both writing and reading.

See Learning, recall and graceful forgetting for the measured behaviour and Read endpoints for read options.