mirror of
https://github.com/samvallad33/vestige.git
synced 2026-04-25 00:36:22 +02:00
f9c60eb5a7
FSRS-6 spaced repetition, spreading activation, synaptic tagging, hippocampal indexing, and 130 years of memory research. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
606 lines
21 KiB
Rust
606 lines
21 KiB
Rust
//! # Speculative Memory Retrieval
|
|
//!
|
|
//! Predict what memories the user will need BEFORE they ask.
|
|
//! Uses pattern analysis, temporal modeling, and context understanding
|
|
//! to pre-warm the cache with likely-needed memories.
|
|
//!
|
|
//! ## How It Works
|
|
//!
|
|
//! 1. Analyzes current working context (files open, recent queries, project state)
|
|
//! 2. Learns from historical access patterns (what memories were accessed together)
|
|
//! 3. Predicts with confidence scores and reasoning
|
|
//! 4. Pre-fetches high-confidence predictions into fast cache
|
|
//! 5. Records actual usage to improve future predictions
|
|
//!
|
|
//! ## Example
|
|
//!
|
|
//! ```rust,ignore
|
|
//! let retriever = SpeculativeRetriever::new(storage);
|
|
//!
|
|
//! // When user opens auth.rs, predict they'll need JWT memories
|
|
//! let predictions = retriever.predict_needed(&context);
|
|
//!
|
|
//! // Pre-warm cache in background
|
|
//! retriever.prefetch(&context).await?;
|
|
//! ```
|
|
|
|
use chrono::{DateTime, Timelike, Utc};
|
|
use serde::{Deserialize, Serialize};
|
|
use std::collections::{HashMap, VecDeque};
|
|
use std::path::PathBuf;
|
|
use std::sync::{Arc, RwLock};
|
|
|
|
/// Maximum number of access patterns to track
|
|
const MAX_PATTERN_HISTORY: usize = 10_000;
|
|
|
|
/// Maximum predictions to return
|
|
const MAX_PREDICTIONS: usize = 20;
|
|
|
|
/// Minimum confidence threshold for predictions
|
|
const MIN_CONFIDENCE: f64 = 0.3;
|
|
|
|
/// Decay factor for old patterns (per day)
|
|
const PATTERN_DECAY_RATE: f64 = 0.95;
|
|
|
|
/// A predicted memory that the user is likely to need
|
|
#[derive(Debug, Clone, Serialize, Deserialize)]
|
|
pub struct PredictedMemory {
|
|
/// The memory ID that's predicted to be needed
|
|
pub memory_id: String,
|
|
/// Content preview for quick reference
|
|
pub content_preview: String,
|
|
/// Confidence score (0.0 to 1.0)
|
|
pub confidence: f64,
|
|
/// Human-readable reasoning for this prediction
|
|
pub reasoning: String,
|
|
/// What triggered this prediction
|
|
pub trigger: PredictionTrigger,
|
|
/// When this prediction was made
|
|
pub predicted_at: DateTime<Utc>,
|
|
}
|
|
|
|
/// What triggered a prediction
|
|
#[derive(Debug, Clone, Serialize, Deserialize)]
|
|
pub enum PredictionTrigger {
|
|
/// Based on file being opened/edited
|
|
FileContext { file_path: String },
|
|
/// Based on co-access patterns
|
|
CoAccessPattern { related_memory_id: String },
|
|
/// Based on time-of-day patterns
|
|
TemporalPattern { typical_time: String },
|
|
/// Based on project context
|
|
ProjectContext { project_name: String },
|
|
/// Based on detected intent
|
|
IntentBased { intent: String },
|
|
/// Based on semantic similarity to recent queries
|
|
SemanticSimilarity { query: String, similarity: f64 },
|
|
}
|
|
|
|
/// Context for making predictions
|
|
#[derive(Debug, Clone, Default)]
|
|
pub struct PredictionContext {
|
|
/// Currently open files
|
|
pub open_files: Vec<PathBuf>,
|
|
/// Recent file edits
|
|
pub recent_edits: Vec<PathBuf>,
|
|
/// Recent search queries
|
|
pub recent_queries: Vec<String>,
|
|
/// Recently accessed memory IDs
|
|
pub recent_memory_ids: Vec<String>,
|
|
/// Current project path
|
|
pub project_path: Option<PathBuf>,
|
|
/// Current timestamp
|
|
pub timestamp: Option<DateTime<Utc>>,
|
|
}
|
|
|
|
impl PredictionContext {
|
|
/// Create a new prediction context
|
|
pub fn new() -> Self {
|
|
Self {
|
|
timestamp: Some(Utc::now()),
|
|
..Default::default()
|
|
}
|
|
}
|
|
|
|
/// Add an open file to context
|
|
pub fn with_file(mut self, path: PathBuf) -> Self {
|
|
self.open_files.push(path);
|
|
self
|
|
}
|
|
|
|
/// Add a recent query to context
|
|
pub fn with_query(mut self, query: String) -> Self {
|
|
self.recent_queries.push(query);
|
|
self
|
|
}
|
|
|
|
/// Set the project path
|
|
pub fn with_project(mut self, path: PathBuf) -> Self {
|
|
self.project_path = Some(path);
|
|
self
|
|
}
|
|
}
|
|
|
|
/// A learned co-access pattern
|
|
#[derive(Debug, Clone, Serialize, Deserialize)]
|
|
pub struct UsagePattern {
|
|
/// The trigger memory ID
|
|
pub trigger_id: String,
|
|
/// The predicted memory ID
|
|
pub predicted_id: String,
|
|
/// How often this pattern occurred
|
|
pub frequency: u32,
|
|
/// Success rate (was the prediction useful)
|
|
pub success_rate: f64,
|
|
/// Last time this pattern was observed
|
|
pub last_seen: DateTime<Utc>,
|
|
/// Weight after decay applied
|
|
pub weight: f64,
|
|
}
|
|
|
|
/// Speculative memory retriever that predicts needed memories
|
|
pub struct SpeculativeRetriever {
|
|
/// Co-access patterns: trigger_id -> Vec<(predicted_id, pattern)>
|
|
co_access_patterns: Arc<RwLock<HashMap<String, Vec<UsagePattern>>>>,
|
|
/// File-to-memory associations
|
|
file_memory_map: Arc<RwLock<HashMap<String, Vec<String>>>>,
|
|
/// Recent access sequence for pattern detection
|
|
access_sequence: Arc<RwLock<VecDeque<AccessEvent>>>,
|
|
/// Pending predictions (for recording outcomes)
|
|
pending_predictions: Arc<RwLock<HashMap<String, PredictedMemory>>>,
|
|
/// Cache of recently predicted memories
|
|
prediction_cache: Arc<RwLock<Vec<PredictedMemory>>>,
|
|
}
|
|
|
|
/// An access event for pattern learning
|
|
#[derive(Debug, Clone, Serialize, Deserialize)]
|
|
struct AccessEvent {
|
|
memory_id: String,
|
|
file_context: Option<String>,
|
|
query_context: Option<String>,
|
|
timestamp: DateTime<Utc>,
|
|
was_helpful: Option<bool>,
|
|
}
|
|
|
|
impl SpeculativeRetriever {
|
|
/// Create a new speculative retriever
|
|
pub fn new() -> Self {
|
|
Self {
|
|
co_access_patterns: Arc::new(RwLock::new(HashMap::new())),
|
|
file_memory_map: Arc::new(RwLock::new(HashMap::new())),
|
|
access_sequence: Arc::new(RwLock::new(VecDeque::with_capacity(MAX_PATTERN_HISTORY))),
|
|
pending_predictions: Arc::new(RwLock::new(HashMap::new())),
|
|
prediction_cache: Arc::new(RwLock::new(Vec::new())),
|
|
}
|
|
}
|
|
|
|
/// Predict memories that will be needed based on context
|
|
pub fn predict_needed(&self, context: &PredictionContext) -> Vec<PredictedMemory> {
|
|
let mut predictions: Vec<PredictedMemory> = Vec::new();
|
|
let now = context.timestamp.unwrap_or_else(Utc::now);
|
|
|
|
// 1. File-based predictions
|
|
predictions.extend(self.predict_from_files(context, now));
|
|
|
|
// 2. Co-access pattern predictions
|
|
predictions.extend(self.predict_from_patterns(context, now));
|
|
|
|
// 3. Query similarity predictions
|
|
predictions.extend(self.predict_from_queries(context, now));
|
|
|
|
// 4. Temporal pattern predictions
|
|
predictions.extend(self.predict_from_time(now));
|
|
|
|
// Deduplicate and sort by confidence
|
|
predictions = self.deduplicate_predictions(predictions);
|
|
predictions.sort_by(|a, b| b.confidence.partial_cmp(&a.confidence).unwrap_or(std::cmp::Ordering::Equal));
|
|
predictions.truncate(MAX_PREDICTIONS);
|
|
|
|
// Filter by minimum confidence
|
|
predictions.retain(|p| p.confidence >= MIN_CONFIDENCE);
|
|
|
|
// Store for outcome tracking
|
|
self.store_pending_predictions(&predictions);
|
|
|
|
predictions
|
|
}
|
|
|
|
/// Pre-warm cache with predicted memories
|
|
pub async fn prefetch(&self, context: &PredictionContext) -> Result<usize, SpeculativeError> {
|
|
let predictions = self.predict_needed(context);
|
|
let count = predictions.len();
|
|
|
|
// Store predictions in cache for fast access
|
|
if let Ok(mut cache) = self.prediction_cache.write() {
|
|
*cache = predictions;
|
|
}
|
|
|
|
Ok(count)
|
|
}
|
|
|
|
/// Record what was actually used to improve future predictions
|
|
pub fn record_usage(&self, _predicted: &[String], actually_used: &[String]) {
|
|
// Update pending predictions with outcomes
|
|
if let Ok(mut pending) = self.pending_predictions.write() {
|
|
for id in actually_used {
|
|
if let Some(prediction) = pending.remove(id) {
|
|
// This was correctly predicted - strengthen pattern
|
|
self.strengthen_pattern(&prediction.memory_id, 1.0);
|
|
}
|
|
}
|
|
|
|
// Weaken patterns for predictions that weren't used
|
|
for (id, _) in pending.drain() {
|
|
self.weaken_pattern(&id, 0.9);
|
|
}
|
|
}
|
|
|
|
// Learn new co-access patterns
|
|
self.learn_co_access_patterns(actually_used);
|
|
}
|
|
|
|
/// Record a memory access event
|
|
pub fn record_access(
|
|
&self,
|
|
memory_id: &str,
|
|
file_context: Option<&str>,
|
|
query_context: Option<&str>,
|
|
was_helpful: Option<bool>,
|
|
) {
|
|
let event = AccessEvent {
|
|
memory_id: memory_id.to_string(),
|
|
file_context: file_context.map(String::from),
|
|
query_context: query_context.map(String::from),
|
|
timestamp: Utc::now(),
|
|
was_helpful,
|
|
};
|
|
|
|
if let Ok(mut sequence) = self.access_sequence.write() {
|
|
sequence.push_back(event.clone());
|
|
|
|
// Trim old events
|
|
while sequence.len() > MAX_PATTERN_HISTORY {
|
|
sequence.pop_front();
|
|
}
|
|
}
|
|
|
|
// Update file-memory associations
|
|
if let Some(file) = file_context {
|
|
if let Ok(mut map) = self.file_memory_map.write() {
|
|
map.entry(file.to_string())
|
|
.or_insert_with(Vec::new)
|
|
.push(memory_id.to_string());
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Get cached predictions
|
|
pub fn get_cached_predictions(&self) -> Vec<PredictedMemory> {
|
|
self.prediction_cache
|
|
.read()
|
|
.map(|cache| cache.clone())
|
|
.unwrap_or_default()
|
|
}
|
|
|
|
/// Apply decay to old patterns
|
|
pub fn apply_pattern_decay(&self) {
|
|
if let Ok(mut patterns) = self.co_access_patterns.write() {
|
|
let now = Utc::now();
|
|
|
|
for patterns_list in patterns.values_mut() {
|
|
for pattern in patterns_list.iter_mut() {
|
|
let days_old = (now - pattern.last_seen).num_days() as f64;
|
|
pattern.weight = pattern.weight * PATTERN_DECAY_RATE.powf(days_old);
|
|
}
|
|
|
|
// Remove patterns that are too weak
|
|
patterns_list.retain(|p| p.weight > 0.01);
|
|
}
|
|
}
|
|
}
|
|
|
|
// ========================================================================
|
|
// Private prediction methods
|
|
// ========================================================================
|
|
|
|
fn predict_from_files(
|
|
&self,
|
|
context: &PredictionContext,
|
|
now: DateTime<Utc>,
|
|
) -> Vec<PredictedMemory> {
|
|
let mut predictions = Vec::new();
|
|
|
|
if let Ok(file_map) = self.file_memory_map.read() {
|
|
for file in &context.open_files {
|
|
let file_str = file.to_string_lossy().to_string();
|
|
if let Some(memory_ids) = file_map.get(&file_str) {
|
|
for memory_id in memory_ids {
|
|
predictions.push(PredictedMemory {
|
|
memory_id: memory_id.clone(),
|
|
content_preview: String::new(), // Would be filled by storage lookup
|
|
confidence: 0.7,
|
|
reasoning: format!(
|
|
"You're working on {}, and this memory was useful for that file before",
|
|
file.file_name().unwrap_or_default().to_string_lossy()
|
|
),
|
|
trigger: PredictionTrigger::FileContext {
|
|
file_path: file_str.clone()
|
|
},
|
|
predicted_at: now,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
predictions
|
|
}
|
|
|
|
fn predict_from_patterns(
|
|
&self,
|
|
context: &PredictionContext,
|
|
now: DateTime<Utc>,
|
|
) -> Vec<PredictedMemory> {
|
|
let mut predictions = Vec::new();
|
|
|
|
if let Ok(patterns) = self.co_access_patterns.read() {
|
|
for recent_id in &context.recent_memory_ids {
|
|
if let Some(related_patterns) = patterns.get(recent_id) {
|
|
for pattern in related_patterns {
|
|
let confidence = pattern.weight * pattern.success_rate;
|
|
if confidence >= MIN_CONFIDENCE {
|
|
predictions.push(PredictedMemory {
|
|
memory_id: pattern.predicted_id.clone(),
|
|
content_preview: String::new(),
|
|
confidence,
|
|
reasoning: format!(
|
|
"You accessed a related memory, and these are often used together ({}% of the time)",
|
|
(pattern.success_rate * 100.0) as u32
|
|
),
|
|
trigger: PredictionTrigger::CoAccessPattern {
|
|
related_memory_id: recent_id.clone()
|
|
},
|
|
predicted_at: now,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
predictions
|
|
}
|
|
|
|
fn predict_from_queries(
|
|
&self,
|
|
context: &PredictionContext,
|
|
now: DateTime<Utc>,
|
|
) -> Vec<PredictedMemory> {
|
|
// In a full implementation, this would use semantic similarity
|
|
// to find memories similar to recent queries
|
|
let mut predictions = Vec::new();
|
|
|
|
if let Ok(sequence) = self.access_sequence.read() {
|
|
for query in &context.recent_queries {
|
|
// Find memories accessed after similar queries
|
|
for event in sequence.iter().rev().take(100) {
|
|
if let Some(event_query) = &event.query_context {
|
|
// Simple substring matching (would use embeddings in production)
|
|
if event_query.to_lowercase().contains(&query.to_lowercase())
|
|
|| query.to_lowercase().contains(&event_query.to_lowercase())
|
|
{
|
|
predictions.push(PredictedMemory {
|
|
memory_id: event.memory_id.clone(),
|
|
content_preview: String::new(),
|
|
confidence: 0.6,
|
|
reasoning: format!(
|
|
"This memory was helpful when you searched for similar terms before"
|
|
),
|
|
trigger: PredictionTrigger::SemanticSimilarity {
|
|
query: query.clone(),
|
|
similarity: 0.8,
|
|
},
|
|
predicted_at: now,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
predictions
|
|
}
|
|
|
|
fn predict_from_time(&self, now: DateTime<Utc>) -> Vec<PredictedMemory> {
|
|
let mut predictions = Vec::new();
|
|
let hour = now.hour();
|
|
|
|
if let Ok(sequence) = self.access_sequence.read() {
|
|
// Find memories frequently accessed at this time of day
|
|
let mut time_counts: HashMap<String, u32> = HashMap::new();
|
|
|
|
for event in sequence.iter() {
|
|
if (event.timestamp.hour() as i32 - hour as i32).abs() <= 1 {
|
|
*time_counts.entry(event.memory_id.clone()).or_insert(0) += 1;
|
|
}
|
|
}
|
|
|
|
for (memory_id, count) in time_counts {
|
|
if count >= 3 {
|
|
let confidence = (count as f64 / 10.0).min(0.5);
|
|
predictions.push(PredictedMemory {
|
|
memory_id,
|
|
content_preview: String::new(),
|
|
confidence,
|
|
reasoning: format!("You often access this memory around {}:00", hour),
|
|
trigger: PredictionTrigger::TemporalPattern {
|
|
typical_time: format!("{}:00", hour),
|
|
},
|
|
predicted_at: now,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
|
|
predictions
|
|
}
|
|
|
|
fn deduplicate_predictions(&self, predictions: Vec<PredictedMemory>) -> Vec<PredictedMemory> {
|
|
let mut seen: HashMap<String, PredictedMemory> = HashMap::new();
|
|
|
|
for pred in predictions {
|
|
seen.entry(pred.memory_id.clone())
|
|
.and_modify(|existing| {
|
|
// Keep the one with higher confidence
|
|
if pred.confidence > existing.confidence {
|
|
*existing = pred.clone();
|
|
}
|
|
})
|
|
.or_insert(pred);
|
|
}
|
|
|
|
seen.into_values().collect()
|
|
}
|
|
|
|
fn store_pending_predictions(&self, predictions: &[PredictedMemory]) {
|
|
if let Ok(mut pending) = self.pending_predictions.write() {
|
|
pending.clear();
|
|
for pred in predictions {
|
|
pending.insert(pred.memory_id.clone(), pred.clone());
|
|
}
|
|
}
|
|
}
|
|
|
|
fn strengthen_pattern(&self, memory_id: &str, factor: f64) {
|
|
if let Ok(mut patterns) = self.co_access_patterns.write() {
|
|
for patterns_list in patterns.values_mut() {
|
|
for pattern in patterns_list.iter_mut() {
|
|
if pattern.predicted_id == memory_id {
|
|
pattern.weight = (pattern.weight * factor).min(1.0);
|
|
pattern.frequency += 1;
|
|
pattern.success_rate = (pattern.success_rate * 0.9) + 0.1;
|
|
pattern.last_seen = Utc::now();
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
fn weaken_pattern(&self, memory_id: &str, factor: f64) {
|
|
if let Ok(mut patterns) = self.co_access_patterns.write() {
|
|
for patterns_list in patterns.values_mut() {
|
|
for pattern in patterns_list.iter_mut() {
|
|
if pattern.predicted_id == memory_id {
|
|
pattern.weight *= factor;
|
|
pattern.success_rate = pattern.success_rate * 0.95;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
fn learn_co_access_patterns(&self, memory_ids: &[String]) {
|
|
if memory_ids.len() < 2 {
|
|
return;
|
|
}
|
|
|
|
if let Ok(mut patterns) = self.co_access_patterns.write() {
|
|
// Create patterns between each pair of memories
|
|
for i in 0..memory_ids.len() {
|
|
for j in 0..memory_ids.len() {
|
|
if i != j {
|
|
let trigger = &memory_ids[i];
|
|
let predicted = &memory_ids[j];
|
|
|
|
let patterns_list =
|
|
patterns.entry(trigger.clone()).or_insert_with(Vec::new);
|
|
|
|
if let Some(existing) = patterns_list
|
|
.iter_mut()
|
|
.find(|p| p.predicted_id == *predicted)
|
|
{
|
|
existing.frequency += 1;
|
|
existing.weight = (existing.weight + 0.1).min(1.0);
|
|
existing.last_seen = Utc::now();
|
|
} else {
|
|
patterns_list.push(UsagePattern {
|
|
trigger_id: trigger.clone(),
|
|
predicted_id: predicted.clone(),
|
|
frequency: 1,
|
|
success_rate: 0.5,
|
|
last_seen: Utc::now(),
|
|
weight: 0.5,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
impl Default for SpeculativeRetriever {
|
|
fn default() -> Self {
|
|
Self::new()
|
|
}
|
|
}
|
|
|
|
/// Errors that can occur during speculative retrieval
|
|
#[derive(Debug, thiserror::Error)]
|
|
pub enum SpeculativeError {
|
|
/// Failed to access pattern data
|
|
#[error("Pattern access error: {0}")]
|
|
PatternAccess(String),
|
|
|
|
/// Failed to prefetch memories
|
|
#[error("Prefetch error: {0}")]
|
|
Prefetch(String),
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
|
|
#[test]
|
|
fn test_prediction_context() {
|
|
let context = PredictionContext::new()
|
|
.with_file(PathBuf::from("/src/auth.rs"))
|
|
.with_query("JWT token".to_string())
|
|
.with_project(PathBuf::from("/my/project"));
|
|
|
|
assert_eq!(context.open_files.len(), 1);
|
|
assert_eq!(context.recent_queries.len(), 1);
|
|
assert!(context.project_path.is_some());
|
|
}
|
|
|
|
#[test]
|
|
fn test_record_access() {
|
|
let retriever = SpeculativeRetriever::new();
|
|
|
|
retriever.record_access(
|
|
"mem-123",
|
|
Some("/src/auth.rs"),
|
|
Some("JWT token"),
|
|
Some(true),
|
|
);
|
|
|
|
// Verify file-memory association was recorded
|
|
let map = retriever.file_memory_map.read().unwrap();
|
|
assert!(map.contains_key("/src/auth.rs"));
|
|
}
|
|
|
|
#[test]
|
|
fn test_learn_co_access_patterns() {
|
|
let retriever = SpeculativeRetriever::new();
|
|
|
|
retriever.learn_co_access_patterns(&[
|
|
"mem-1".to_string(),
|
|
"mem-2".to_string(),
|
|
"mem-3".to_string(),
|
|
]);
|
|
|
|
let patterns = retriever.co_access_patterns.read().unwrap();
|
|
assert!(patterns.contains_key("mem-1"));
|
|
assert!(patterns.contains_key("mem-2"));
|
|
}
|
|
}
|