---
id: f3790fff-ab75-44b2-8d8b-7dd0953c05dc
type: solution
title: "Embeddings mtime-based cache: 6x faster semantic recall"
tags: [cognitive-memory, python, performance, caching]
importance: 0.8
confidence: 0.8
created: "2026-02-19T22:03:16.251366+00:00"
updated: "2026-02-19T22:03:16.251366+00:00"
---

Added mtime-based caching for `_embeddings.json` (24MB, 439 entries × 4096-dim vectors) in CognitiveMemoryClient.

**Problem:** Every `semantic_recall()` call re-parsed the 24MB JSON file from disk.

**Solution:** Added `_embeddings_cache` and `_embeddings_mtime` instance attributes. New `_load_embeddings_cached()` method does `stat()` to check mtime (nearly free), only re-parses when file has changed. Since the embed cron runs hourly, the parse happens at most once per hour.

**Performance results:**
| Call | Before | After |
|------|--------|-------|
| Semantic (cold) | 1,328ms | 389ms |
| Semantic (warm) | 1,328ms | 208ms |
| Keyword only | 3ms | 2ms |

The remaining ~200ms on warm cache is the Ollama embedding API call for the query text + cosine similarity computation. The JSON parse overhead is completely eliminated on repeat calls.

**Files changed:** `client.py` — added `_load_embeddings_cached()`, updated `semantic_recall()` to use it, added cache attrs to `__init__`.