What Is the Cost of Running Memory Consolidation

Cost Components

Clustering and Similarity Computation

The first phase of consolidation groups related memories by computing entity overlap and vector similarity between pairs. For a store of N memories, this involves at most N*(N-1)/2 pairwise comparisons, but in practice the comparison space is pruned by only comparing memories that share at least one entity. This reduces the computation to a fraction of the theoretical maximum. For 10,000 memories, the clustering phase typically completes in seconds on standard hardware and has negligible cost.

LLM Calls for Contradiction Detection

Contradiction detection is the most expensive component because it requires an LLM to analyze pairs of memories for conflicting factual claims. Not every pair requires an LLM call, only pairs within clusters that have high entity overlap and semantic similarity, where contradiction is plausible. For a store of 10,000 memories with typical redundancy levels, a consolidation run might evaluate 100 to 300 pairs for contradictions. At $0.01 to $0.03 per LLM call (depending on the model used and token counts), this costs $1.00 to $9.00 per run.

You can reduce this cost by using a cheaper NLI model for initial screening and only escalating ambiguous cases to an LLM. A DeBERTa-based NLI model running on a GPU instance can classify pairs as entailment, contradiction, or neutral at a fraction of a cent per pair, filtering out the 80% of pairs that are clearly not contradictions before the expensive LLM analysis.

Embedding Regeneration

Every merged memory needs a new vector embedding because its content has changed. Embedding costs depend on the model and provider, but typical costs are around $0.0001 per embedding for a 1,536-dimension model. If a consolidation run produces 200 merged memories, the embedding cost is about $0.02. This component is essentially negligible compared to the LLM costs.

Entity Re-extraction

Merged memories may need updated entity extraction if the merge changed which entities are referenced. If entity extraction uses an LLM, this adds a per-merge cost similar to the contradiction detection calls. If it uses a specialized NER model, the cost is minimal. A typical run that produces 200 merges with LLM-based entity extraction adds $2.00 to $6.00.

Total Cost Estimates

For a store of 10,000 memories with weekly consolidation, expect $2 to $10 per run, or roughly $10 to $50 per month. For a store of 50,000 memories with nightly consolidation, expect $5 to $25 per run, or $150 to $750 per month. For incremental consolidation that only processes new memories since the last run, costs are proportional to the number of new memories rather than the total store size, which makes incremental runs significantly cheaper.

Cost Recovery

Consolidation costs are an investment that pays back through reduced ongoing expenses. A 10,000-memory store without consolidation grows by the full ingestion rate every month, accumulating storage and compute costs linearly. With consolidation reducing the effective store size by 30% to 40%, the ongoing savings in vector storage, retrieval compute, and embedding storage quickly exceed the consolidation costs.

For most systems, consolidation pays for itself within the first month. A store that would have grown to 15,000 memories without consolidation but stabilizes at 10,000 with it saves 5,000 vectors worth of storage and retrieval costs every month, while the consolidation runs themselves cost a fraction of those savings.

Optimizing Consolidation Costs

Use incremental consolidation for regular runs, processing only new memories since the last cycle. Reserve full consolidation for monthly or quarterly runs. Use cheaper NLI models for contradiction screening and escalate only ambiguous cases to an LLM. Set a maximum number of LLM calls per run to cap costs. Process memories in batches so you can stop a run at a cost boundary without losing the work already completed.

Adaptive Recall handles these optimizations internally. The reflect tool uses incremental processing by default, batches LLM calls efficiently, and reports the cost of each consolidation run through the status tool so you can monitor spending.