Learning, recall and graceful forgetting
HeatherDB's competitive mode supports associative recall and shared-pattern learning. Plain mode preserves individual vector locations. The two modes use different write representations.
How writes represent information
In competitive mode, a novel pattern receives a location of its own without changing existing counters or addresses. A joining write accumulates content and moves addresses within the activated representation. Related observations can therefore reinforce shared structure, while distinct patterns can remain separately represented.
Plain writes append each vector as an individual location without those adaptive updates. Reads can still select or combine these locations through attention. Neither representation guarantees exact reconstruction for arbitrary cues, occupancy or data distributions.
Measure recall and learning separately
Recall compares a reconstruction with the particular pattern its cue came from. Shared-pattern learning measures whether new noisy observations improve responses on independent examples of the underlying structure. Both objectives can be relevant to competitive memory. One metric does not substitute for the other, and low recall alone is not evidence of useful abstraction.
A seeded recall fixture uses 32 random unit patterns in 64 dimensions and erases every fourth coordinate of each cue:
| Write mode and readout | Mean cosine to original | Fraction at cosine ≥ 0.9 |
|---|---|---|
| Competitive, default β=5, iterative | 0.8844 | 59.375% |
| Competitive, explicit β=50, iterative | 0.8977 | 65.625% |
| Plain, attention scale=50 | 1.0000 | 100.00% |
These are configuration-specific reconstruction measurements, not a universal capacity bound or a recommendation to change the default temperature. The acceptance evidence records the source revision, fixture and minimum scores.
In a separate learning task with eight latent patterns, 128 dimensions and five data seeds, default competitive mode reaches mean held-out prototype cosine 0.9941 after 1,024 noisy observations. The noisy input scores 0.8009 and an online k-means baseline with the same location count scores 0.9968. This establishes learning on this synthetic task, without establishing superiority to clustering or semantic reasoning. See the full evaluation.
When patterns change
An adaptive memory can increasingly reconstruct a changed pattern as new observations arrive. Its response may consequently become less similar to the earlier pattern. For a policy that prioritises current knowledge, this is graceful forgetting. Assess it through improved response to the new patterns alongside retention of the old ones. Unwanted interference between unrelated patterns is not, by itself, evidence of graceful forgetting.
A synthetic drift task learns eight latent patterns from noisy observations, then rotates them by 45°. These five-seed means use independent held-out queries and a common attention scale of 50:
| New-pattern observations | Fidelity to new patterns | Fidelity to old patterns |
|---|---|---|
| 0 | 0.6939 | 0.9799 |
| 64 | 0.8844 | 0.9250 |
| 256 | 0.9848 | 0.7838 |
| 1,024 | 0.9920 | 0.7109 |
Fidelity is cosine to the corresponding clean latent pattern. The measured checkpoints show adaptation and decreasing old-pattern influence. They do not establish a fixed forgetting rate, monotonic behaviour on all data or application-level accuracy. Protocol, baselines and raw results are in the learning evaluation.
Retention and document identity
A reconstruction can express a learned pattern without identifying one original observation. Metadata-bearing writes also retain document records; use the document API when your application needs exact record identity. Use a separate plain database when it also needs independently retained vector locations. See Choose a memory mode.
Graceful forgetting is not a TTL, guaranteed age-based decay or secure unlearning.
lambda decays the learning rate, not observation lifetimes. Deleting a document
does not subtract its learned influence from shared counters, and old influence
may remain even after responses adapt.
Discovery requires validation
Competitive learning can supply candidate patterns for investigation. A learned direction does not establish a distinct, meaningful or new discovery. Evaluate candidates on independent data against appropriate nulls and simple baselines. The synthetic study uses external validation; native candidate generation does not provide automatic validation or reliable unknown-query rejection.